Safeguard
AI Security•

What is artificial intelligence? History, types, how AI is built and governed

What is artificial intelligence? Definitions from Turing to the EU AI Act, AI history, narrow vs general AI, how AI systems are built, and AI regulation.

Safeguard Security Team
39 min read

Artificial intelligence (AI) is the field of building machine-based systems that take inputs and infer outputs such as predictions, content, recommendations or decisions, with some degree of autonomy. That wording is close to the legal definition in Article 3(1) of the EU AI Act and to the OECD definition revised in November 2023. Deployed AI systems are very good at bounded sets of tasks, and no system has been shown to meet the published bar for Competent general intelligence.

This hub post covers what AI means in law and research, where it came from, the main families of techniques, how a modern AI system is built, and how it is governed in November 2026, plus why security teams should treat AI as a supply chain rather than a single product.

Key takeaways

  • The legal definitions have converged. The EU AI Act and the OECD both describe a machine-based system that infers outputs from inputs.
  • AI is older than most people think. The term comes from a proposal dated 31 August 1955, decades before deep learning.
  • General intelligence has not been shown. The Levels of AGI framework from Google DeepMind researchers rates 2023 frontier chatbots as "Emerging" and says no public system had reached "Competent AGI" at the time of writing (latest revision September 2025).
  • Modern AI is a pipeline: data, pretraining, post-training, evaluation, serving, retrieval, tools and monitoring. Each stage leaves artefacts you have to secure.
  • The EU rules moved in 2026. Regulation (EU) 2026/1744, the "Digital Omnibus on AI", published 24 July 2026, set the high-risk dates at 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
  • US national AI policy favours one framework. Executive Order 14365 (11 December 2025) set up a task force to challenge conflicting state AI laws.

This is part 1 of a four-part series on AI foundations.

  1. What is artificial intelligence? History, types, how AI is built and governed (this post)
  2. What is machine learning? Supervised, unsupervised and reinforcement learning, models, MLOps and attacks
  3. What is deep learning? Neural networks, transformers, diffusion, scaling and model security
  4. What is superintelligence? AGI vs ASI, timelines, alignment research and governance

What is artificial intelligence? Five definitions that matter

People use "AI" for anything from a spam filter to a science fiction mind. These definitions carry weight in engineering and compliance work.

Turing, 1950: change the question

Alan Turing's paper "Computing Machinery and Intelligence" appeared in the journal Mind in October 1950 (volume LIX, pages 433 to 460). Turing replaced the vague question "can machines think?" with a behavioural test: can a machine hold a text conversation well enough that an interrogator cannot tell it from a person? It moved the debate from what a machine is to what it does, and the definitions below follow that move.

Dartmouth, 1955: name the field

The phrase "artificial intelligence" comes from "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence", dated 31 August 1955 and signed by John McCarthy (Dartmouth College), Marvin Minsky (Harvard University), Nathaniel Rochester (IBM) and Claude Shannon (Bell Telephone Laboratories). It asked for a two month, ten person study in the summer of 1956, built on the conjecture that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." The target problems included using language, forming abstractions and concepts, solving problems then reserved for humans, and self-improvement.

The OECD, November 2023

The OECD definition matters because many national laws copy it. On 29 November 2023, Stuart Russell, Karine Perset and Marko Grobelnik explained the update OECD members had agreed: an AI system is a machine-based system that, for explicit or implicit objectives, infers from its input how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments, with varying levels of autonomy and adaptiveness. The edits added "content" as an output, a nod to generative models, and replaced "real" with "physical" environments.

The EU AI Act, Article 3(1)

Regulation (EU) 2024/1689, the AI Act, defines the term that decides whether the law applies to you:

'AI system' means a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.

The word doing the work is infers. A rules engine that only executes rules a person wrote line by line is a weak fit. A model that learned its behaviour from data and generalises to new inputs is a clear fit. The AI Act Service Desk links the Commission's guidelines on the definition, worth reading before you decide a system is out of scope.

ISO/IEC 22989:2022

ISO/IEC 22989:2022, "Artificial intelligence concepts and terminology", was published in 2022 by ISO/IEC JTC 1/SC 42. It is the vocabulary standard the rest of the SC 42 family builds on, and ISO/IEC 42001 sits in the same family. When two teams argue about what a "model" or an "AI agent" means, start here.

A short history of AI, with dates

AI history is a set of competing ideas that took turns being fashionable, with gaps where funding fell away. The timeline sticks to events dated from a primary source.

DateEventWhy it matteredSource
Oct 1950Turing, "Computing Machinery and Intelligence"Behavioural test for machine intelligenceMind LIX
31 Aug 1955Dartmouth proposalNames the field "artificial intelligence"McCarthy et al.
1968Hart, Nilsson and Raphael, A* searchOptimal heuristic search, still used in routing and gamesIEEE Trans. SSC 4
July 1972Lighthill report to the UK Science Research CouncilA sceptical official review of AI progressLighthill
Apr 1975Shortliffe and Buchanan, inexact reasoning in medicineRule-based reasoning under uncertain evidenceMathematical Biosciences 23
Sep 1982McDermott, "R1: a rule-based configurer of computer systems"Rules applied to system configurationArtificial Intelligence 19
Oct 1986Rumelhart, Hinton and Williams, back-propagationPractical training of multi-layer neural networksNature 323
2012Krizhevsky, Sutskever and Hinton (AlexNet)Deep convolutional network wins ILSVRC-2012NIPS 2012
2016Silver et al., AlphaGoDeep networks plus tree search beat a professional Go playerNature 529
Jun 2017Vaswani et al., "Attention Is All You Need"The transformer architecturearXiv 1706.03762
May 2020Lewis et al., retrieval-augmented generationModels that look things up before answeringarXiv 2005.11401
Aug 2021Bommasani et al., foundation modelsNames the "train once, adapt everywhere" patternarXiv 2108.07258
Mar 2022Ouyang et al., InstructGPTPost-training with human feedbackarXiv 2203.02155
Oct 2022Yao et al., ReActLanguage models that reason and call tools in a looparXiv 2210.03629
26 Jan 2023NIST AI RMF 1.0US voluntary AI risk frameworkNIST AI 100-1
18 Dec 2023ISO/IEC 42001:2023First AI management system standardIEC webstore
26 Jul 2024NIST AI 600-1Generative AI profile of the AI RMFNIST
1 Aug 2024EU AI Act enters into forceFirst horizontal AI lawAI Act Service Desk
24 Jul 2026Regulation (EU) 2026/1744 in the Official JournalDigital Omnibus on AI changes AI Act datesEuropean Parliament OEIL

Symbolic AI and search

The first generation of AI treated intelligence as symbol manipulation: represent the world as symbols, write rules over them, and search for a sequence of steps that reaches a goal. The 1968 A* paper by Peter Hart, Nils Nilsson and Bertram Raphael showed how to find a minimum cost path through a graph with a heuristic estimate, and proved the path is optimal if the heuristic never overestimates the remaining cost. We work an example later in this post.

The first downturn

By the early 1970s the gap between promises and results was hard to ignore. The Lighthill report to the UK Science Research Council, dated July 1972, said that in no part of the field had discoveries "produced the major impact that was then promised." It also warned that new predictions of machine intelligence exceeding human intelligence "by the year 2000" should be checked against how earlier predictions had turned out. Later writers use "AI winter" for periods of reduced funding and enthusiasm that followed reports like this one.

Expert systems

One response to the critics was to stop chasing general intelligence and capture narrow expert knowledge as rules. Two papers show the pattern. Edward Shortliffe and Bruce Buchanan described "A model of inexact reasoning in medicine" (April 1975), a way for a rule-based program to reason with uncertain evidence. John McDermott's "R1: A rule-based configurer of computer systems" (September 1982) applied rules to configuring computer systems. Rule bases are easy to audit but only know what someone wrote down, and keeping them current is manual work. Machine learning has its own version of that problem in training data that goes stale.

Statistical machine learning

Instead of writing the rules, you can learn them from examples. Frank Rosenblatt's 1958 perceptron paper is an early landmark, and the 1986 Nature paper by Rumelhart, Hinton and Williams made training networks with hidden layers practical. Part 2 covers the classic methods with worked numbers.

Deep learning

The 2012 paper by Krizhevsky, Sutskever and Hinton described a convolutional network with 60 million parameters. A variant won ILSVRC-2012 with a top-5 test error of 15.3%, against 26.2% for the second-best entry, which convinced much of the field that large networks, large datasets and GPUs were the way forward. In 2016 AlphaGo combined value and policy networks with tree search and, according to the Nature paper, defeated the European Go champion by 5 games to 0.

Transformers and foundation models

The June 2017 transformer paper replaced recurrence with attention and reported 28.4 BLEU on WMT 2014 English to German. In January 2020 Kaplan and colleagues reported that language model loss falls as a smooth power law in model size, dataset size and compute. In August 2021 Bommasani and more than 100 co-authors coined "foundation models" for models "trained on broad data at scale" that are "adaptable to a wide range of downstream tasks." That phrase describes most of what people now call AI.

Narrow AI vs general AI

Narrow AI (sometimes called weak AI or artificial narrow intelligence, ANI) performs well on a defined task or set of tasks. A spam filter, a chess engine, a protein structure predictor and a fraud model are all narrow, however impressive they are. General AI (artificial general intelligence, AGI), in the Levels of AGI paper, would perform at least as well as skilled humans across a wide range of non-physical tasks, including learning new skills. Superintelligence (artificial superintelligence, ASI) would exceed all humans across that range.

Those words get used loosely, so a useful reference is the "Levels of AGI" paper by Meredith Ringel Morris, Shane Legg and colleagues at Google DeepMind (arXiv 2311.02462, first posted 4 November 2023, revised 24 September 2025; the example assignments below are the authors' September 2023 assessment). It separates performance (how good) from generality (how broad) and gives each cell examples.

LevelPerformance barNarrow examples given in the paperGeneral (AGI) status in the paper
0: No AINot AICalculator software, a compilerHuman-in-the-loop computing
1: EmergingEqual to or somewhat better than an unskilled humanRule-based systems such as SHRDLUFrontier chatbots of 2023 (ChatGPT, Bard, Llama 2, Gemini)
2: CompetentAt least the 50th percentile of skilled adultsSmart speakers, toxicity detectorsNot yet achieved
3: ExpertAt least the 90th percentile of skilled adultsSpelling and grammar checkers, image generatorsNot yet achieved
4: Exceptional (formerly "Virtuoso")At least the 99th percentile of skilled adultsDeep Blue, AlphaGoNot yet achieved
5: SuperhumanOutperforms 100% of humansAlphaFold, AlphaZero, StockfishASI, not yet achieved

Three points follow from the table. First, a narrow system can be superhuman while a general one is still "emerging", so "how smart is it" is the wrong question. Note that the paper itself calls 2023 frontier chatbots general but only at the Emerging level. Second, the authors say the Competent AGI level "best corresponds to many prior conceptions of AGI", and the paper notes that OpenAI's labour-replacement threshold (cited as OpenAI, 2018) matches their Exceptional level, so labs do not share one finish line. Third, the authors admit that unambiguous classification needs a standardised benchmark of tasks that does not yet exist. When someone tells you a system "is AGI", ask which definition and which benchmark.

For security work, what a system is allowed to do matters more. The paper adds autonomy levels, from AI as a tool to AI as an agent, because risk changes with autonomy even when capability does not.

The families of AI techniques

"AI" covers several engineering traditions. Most production systems mix them, and each brings a different failure mode.

FamilyCore ideaClassic exampleWhere you meet it todayTypical failure
Search and planningExplore a space of states to reach a goalA* (1968), game-tree searchRoute planning, schedulers, agent plannersBad heuristic, state explosion
Knowledge representation and reasoningEncode facts and rules, derive conclusionsRule-based programs (Shortliffe and Buchanan 1975, McDermott 1982)Policy engines, ontologies, compliance rulesBrittle outside the encoded rules
Machine learningLearn a function from examplesPerceptron (1958), decision tree ensemblesFraud, spam, ranking, forecastingData drift, leakage, poisoned data
Deep learningLearn layered representations with neural networksAlexNet (2012), transformers (2017)Vision, speech, language modelsAdversarial inputs, opaque behaviour
RoboticsPerception and control in the physical worldIndustrial arms, mobile robotsWarehouses, vehicles, dronesSim-to-real gaps, safety interlocks
AgentsPerceive, plan, act with tools in a loopReAct (2022)Coding agents, support agentsPrompt injection, over-broad permissions

Search and planning, worked through

Search is the easiest family to understand with numbers, and it still sits inside modern agents that plan multi-step tasks. Take a small graph with a start node S and a goal G:

  • S to A costs 1, S to B costs 4
  • A to B costs 2, A to G costs 5
  • B to G costs 1

A* ranks each frontier node by f(n) = g(n) + h(n), where g is the cost so far and h is a heuristic estimate of the cost still to go. Use h(S) = 4, h(A) = 3, h(B) = 1, h(G) = 0. The true remaining costs are S: 4, A: 3, B: 1, so the heuristic never overestimates, which is the condition Hart, Nilsson and Raphael showed guarantees an optimal answer.

  1. Expand S (f = 0 + 4 = 4). Add A with g = 1, f = 1 + 3 = 4, and B with g = 4, f = 4 + 1 = 5.
  2. Expand A (f = 4). Reaching B through A costs g = 1 + 2 = 3, so f(B) drops to 3 + 1 = 4. Reaching G through A costs g = 6, f = 6.
  3. Expand B (f = 4). Reaching G through B costs g = 3 + 1 = 4, f = 4, which beats 6.
  4. Expand G (f = 4). Done. The path is S, A, B, G with cost 4.

S, A, G costs 6 and S, B, G costs 5. A* found the cheaper route by updating B when a better path to it appeared.

Game-tree search extends the idea to two players who alternate moves. AlphaGo kept tree search but used learned networks to evaluate positions, a good picture of how the families combine.

Knowledge representation and reasoning

The system stores facts and rules explicitly and derives new facts from them, as policy engines, compliance rule sets and knowledge graphs still do. Every conclusion traces to a readable rule, but the system only knows what someone wrote down. Many products pair a learned model with a rule layer: the model handles messy input and the rules enforce hard limits.

Machine learning and deep learning

Machine learning fits a function to data, and deep learning fits it with many-layered neural networks. These two families produce nearly every system regulators now worry about, and the artefacts security teams must track: datasets, weights, training code and evaluation results. Parts 2 and 3 go deep on both.

Agents

An AI agent perceives its environment, picks an action, takes it through a tool, observes the result and repeats. The October 2022 ReAct paper by Shunyu Yao and colleagues showed language models interleaving reasoning steps with actions such as search queries. The Model Context Protocol (MCP), whose 18 June 2025 specification describes "an open protocol" connecting LLM applications to "external data sources and tools", is one common way to give agents tools. For how to evaluate agents, see our post on AI agent evaluation, and for their specific risks, zero-day risk in AI agents, MCP servers and skills.

How a modern AI system is built

A large language model product is not one artefact. It is a pipeline, and each stage leaves something behind that you may have to inventory, secure or explain to an auditor.

StageWhat happensWhat it producesSecurity and governance questions
Data collectionCrawl, license, buy or generate text, images, codeRaw datasets, licences, provenance notesDo you have rights to it? Could it be poisoned?
Curation and tokenisationFilter, deduplicate, remove personal data, split into tokensCleaned corpora, tokenizer filesWhat was removed and why? Is it reproducible?
PretrainingTrain a large model to predict the next token over the corpusBase model weights, checkpointsWhere are checkpoints stored, who can read them?
Post-trainingSupervised fine-tuning, preference tuning (for example RLHF)Instruction-tuned weights, reward modelsWhich data shaped behaviour? Are evals tied to versions?
EvaluationCapability, safety and regression testsEval reports, model cardsAre results dated and reproducible?
PackagingSerialise weights and configsFiles such as safetensors, PyTorch, GGUF, ONNXIs the format safe to load? Is it signed?
ServingRun inference behind an APIEndpoints, quantised variants, configsWho can call it, with what limits and logs?
RetrievalLook up documents at query timeVector stores, indexes, connectorsCan retrieved text inject instructions?
Tool use and agentsLet the model call functions, APIs, MCP serversTool definitions, permissions, skillsLeast privilege, approval steps, audit trails
MonitoringWatch outputs, costs, drift and abuseLogs, guardrail events, incidentsCan you reconstruct what happened?

Pretraining and the arithmetic of size

Pretraining is where most of the compute goes: the model learns to predict the next token over a very large corpus. Parameter count drives storage. A model with 7 billion parameters in 16-bit floating point (2 bytes each) needs 7,000,000,000 × 2 = 14,000,000,000 bytes, about 14 GB, for weights alone; in 32-bit floats 28 GB; quantised to 4 bits about 3.5 GB plus scale overhead. Training needs several times the weight memory for gradients and optimiser state. Our post on LLM training cost by chip and model size works through those numbers.

Post-training

A pretrained base model continues text. It does not reliably follow instructions. The March 2022 InstructGPT paper by Long Ouyang and colleagues at OpenAI fine-tuned GPT-3 on human-written demonstrations, then trained a reward model on human rankings of outputs and optimised against it with reinforcement learning from human feedback (RLHF). In their human evaluations, outputs from the 1.3 billion parameter InstructGPT model were preferred to outputs from the 175 billion parameter GPT-3, despite having 100 times fewer parameters.

Retrieval

Retrieval-augmented generation (RAG), described by Patrick Lewis and colleagues in May 2020, fetches documents at query time and feeds them to the model. Any document the system can retrieve therefore becomes model input, including one written by an attacker, which is the root of indirect prompt injection.

Tools, agents and the permission question

Once a model can call tools, its outputs become actions. Treat every tool as a permission grant: ask what the agent can reach, whether a human approves changes and whether calls are logged. Our post on governing every LLM call with an AI gateway covers the runtime side.

The AI supply chain

From a security point of view, an AI system is a supply chain with more kinds of parts than a normal application. You inherit risk from every one of them.

ComponentWhere it usually comes fromWhat can go wrongA sensible control
Training and fine-tuning dataPublic datasets, scraped web, vendors, your own recordsPoisoned or mislabelled samples, licence problems, personal dataProvenance records, dataset hashes, review of new sources
Model weightsModel hubs, vendors, your own training runsUnsafe serialisation formats, tampered files, unknown lineagePrefer safe formats, scan before load, pin by hash
ML libraries and runtimesPyPI, conda, container imagesKnown CVEs, dependency confusion, abandoned packagesSBOM, vulnerability scanning, version pinning
Hosted model APIsModel providersData handling terms, outages, silent model updatesContract review, version pinning where offered, logging
Prompts, system instructions, skillsYour teams, marketplacesInstructions that leak data or over-grant toolsCode review, inventory, least privilege
MCP servers and toolsOpen source, vendors, internalOver-broad access, untrusted tool outputAllow-lists, approval steps, gateway policy

Two of these deserve extra attention.

Weights are code, sometimes literally. The Python documentation is blunt: "The pickle module is not secure. Only unpickle data you trust." It warns that crafted pickle data can execute arbitrary code during unpickling. Many older PyTorch checkpoints are pickle files under the hood. Loading an untrusted model file in an unsafe format is the AI equivalent of running an unknown installer. See the Python pickle security model and backdoored model weights for the mechanics and defences.

The value chain is a named risk. NIST's Generative AI Profile (NIST AI 600-1, 26 July 2024) lists "Value Chain and Component Integration" among the risks that are novel to or exacerbated by generative AI. It describes non-transparent or untraceable integration of upstream third-party components, including improperly obtained data and improper supplier vetting across the AI lifecycle. That is supply chain security in plain words. An AI bill of materials (AI-BOM) is the practical answer. See AI-BOMs: extending bill of materials thinking to machine learning and our CycloneDX ML-BOM comparison.

For attacks on the models themselves, such as evasion, poisoning, extraction and membership inference, start with what adversarial machine learning is and the MITRE ATLAS matrix. Part 2 of this series maps them to the machine learning lifecycle.

How AI is governed in November 2026

AI governance is a mix of binding law, voluntary frameworks and auditable management standards. A company selling AI into Europe and to US enterprises can meet all three at once.

The EU AI Act

Regulation (EU) 2024/1689 entered into force on 1 August 2024. It sorts AI by use rather than by technique: some practices are prohibited outright, systems used in sensitive areas are "high-risk" and carry the heaviest obligations, some systems carry transparency duties (for example telling people they are interacting with AI or that content is synthetic), and general-purpose AI (GPAI) models have their own chapter.

The Act applies in stages. On 19 November 2025 the Commission proposed the "Digital Omnibus on AI" (COM(2025) 836) to link the high-risk deadlines to the availability of standards and support tools. According to the European Parliament's Legislative Observatory (procedure 2025/0359(COD)), Parliament adopted its first reading position on 16 June 2026, the Council adopted the act on 29 June 2026, it was signed on 8 July 2026 and it was published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744. The Commission's AI Act Service Desk timeline, which says it reflects the omnibus amendments, now reads as follows.

DateWhat appliesStatus on 15 Nov 2026
1 Aug 2024Entry into forceDone
2 Feb 2025General provisions (definitions, AI literacy) and prohibitionsApplying
2 Aug 2025General-purpose AI model rules; governance bodies set upApplying
2 Aug 2026Most remaining rules, transparency rules (Article 50), enforcement starts at national and EU levelApplying
2 Dec 2026New prohibitions on AI generating non-consensual sexual deepfakes and child sexual abuse material; transition deadline for Article 50(2) for certain systems already on the marketUpcoming
2 Aug 2027Each Member State should have at least one AI regulatory sandbox operatingUpcoming
2 Dec 2027High-risk rules for Annex III systems (for example employment, law enforcement)Upcoming
2 Aug 2028High-risk rules for AI embedded in products covered by Annex IUpcoming

Planning documents that still say high-risk obligations start on 2 August 2026 predate the omnibus. The obligations themselves did not go away. Providers of high-risk systems still need technical documentation, risk management, data governance, logging, human oversight and accuracy and cybersecurity measures when their date arrives. Our posts on EU AI Act high-risk obligations and how the EU AI Act and ISO 42001 interact go article by article. Check the dates there against the table above.

For general-purpose AI models, the Commission published a General-Purpose AI Code of Practice made of three separately authored chapters: Transparency, Copyright, and Safety and Security. Providers can sign it to demonstrate compliance with their GPAI obligations.

NIST AI RMF 1.0 and the Generative AI Profile

The NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) was released on 26 January 2023, as directed by the National Artificial Intelligence Initiative Act of 2020. It is voluntary. It organises AI risk work into four functions: Govern, Map, Measure and Manage. On 26 July 2024 NIST released NIST AI 600-1, the Generative Artificial Intelligence Profile, which applies the framework to risks that are new or worse with generative AI, such as confabulation, information security, and value chain and component integration. On 7 April 2026, NIST released a concept note for an AI RMF Profile on Trustworthy AI in Critical Infrastructure, and its AI RMF page says AI RMF 1.0 is being revised as part of the White House AI Action Plan. For how the AI RMF fits with NIST's cybersecurity guidance, see our post on the NIST AI RMF cybersecurity profile.

ISO/IEC 42001:2023

ISO/IEC 42001:2023, "Artificial intelligence, Management system", was published on 18 December 2023 by ISO/IEC JTC 1/SC 42. It specifies requirements for establishing, implementing, maintaining and continually improving an AI management system inside an organisation. It is a management system standard in the same style as ISO/IEC 27001: it governs the organisation's processes, roles, risk treatment and Annex A controls, not any single model, and conformity can be audited. Our post on ISO 42001 for security teams covers the controls.

US national AI policy

US national AI policy has shifted from safety directives toward a pro-innovation, single national framework. The dated steps:

  • 23 January 2025, EO 14179, "Removing Barriers to American Leadership in Artificial Intelligence". It set the policy "to sustain and enhance America's global AI dominance", ordered an AI Action Plan within 180 days, and directed a review of actions taken under the already revoked EO 14110 of 30 October 2023.
  • July 2025, "America's AI Action Plan", published by the White House under the heading "Winning the Race".
  • 11 December 2025, EO 14365, "Ensuring a National Policy Framework for Artificial Intelligence". It set up an AI Litigation Task Force to challenge state AI laws that conflict with the national policy, called for an evaluation of existing state AI laws, directed a proceeding on a national reporting and disclosure standard for AI models, and asked for a legislative recommendation for one uniform national framework.

None of these orders is a statute. NIST's frameworks remain voluntary, sector regulators still apply existing law, and state AI laws stay in force unless a court or Congress says otherwise.

What security teams should take from this

You do not need to be an AI researcher to secure AI. Treat it the way you already treat software.

For AI inside security products and engineering workflows, which this post does not cover, see AI security tools, AI in cyber security: risks and benefits and how a DevOps team can take advantage of AI.

  1. Inventory AI use: models, model APIs, AI SDKs in code, notebooks, prompts, vector stores, agents and MCP servers, including shadow AI in SaaS.
  2. Record provenance for each model (source, version, hash, licence, format, approver) and each dataset.
  3. Refuse unsafe model formats by default, and scan anything else before it reaches a GPU host.
  4. Patch the ordinary stack. AI services run on packages and images with ordinary CVEs.
  5. Put a policy point in front of model calls for injection, jailbreak attempts and personal data leaving the boundary.
  6. Give agents least privilege, with approval for changes and a log of every tool call.
  7. Map controls to one framework and collect evidence continuously, not just before audits.
  8. Write an AI incident playbook: who can pull a model, revoke a key or disable an agent.

Common mistakes

  • Using a vendor's marketing definition of AI to decide whether a law applies. Use Article 3(1) and the Commission guidelines.
  • Planning against pre-omnibus dates. The Annex III high-risk date is now 2 December 2027 and the Annex I date is 2 August 2028.
  • Treating "the model" as the only AI asset. Datasets, prompts, retrieval indexes, tools and agent permissions all change behaviour and risk.
  • Loading a downloaded checkpoint to "just test it". Unsafe formats can run code on load, so the test is the compromise.
  • Assuming a famous model is general intelligence. On the Levels of AGI scale, no public system had reached Competent AGI at the time of writing, however it is marketed.
  • Treating NIST AI RMF as something you pass. It is a voluntary framework for managing risk. ISO/IEC 42001 is the requirements standard an auditor can assess you against.
  • Forgetting that an agent's risk scales with its permissions, not its intelligence.
  • Reading US state law as dead because of EO 14365. The order starts challenges and proceedings, it does not repeal state statutes.

FAQ

What is artificial intelligence in simple terms?

Software that infers how to produce outputs such as answers, predictions, images or decisions from its inputs, instead of only following rules a person wrote line by line.

What are the main types of AI?

By capability: narrow AI, general AI and superintelligence. Current systems are narrow or, at most, "Emerging" general AI on the Levels of AGI scale; Competent general AI and superintelligence are not achieved. By technique: search and planning, knowledge representation, machine learning, deep learning, robotics and agents.

What is the difference between AI, machine learning and deep learning?

AI is the whole field. Machine learning is the part of AI that learns behaviour from data. Deep learning is the part of machine learning that uses many-layered neural networks. Parts 2 and 3 of this series cover machine learning and deep learning.

Who coined the term artificial intelligence and when?

The Dartmouth Summer Research Project proposal, dated 31 August 1955 and signed by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, proposed a study for the summer of 1956.

Is today's AI general intelligence?

Not at the Competent level or above. The Levels of AGI paper (revised September 2025, with its examples from September 2023) puts frontier chatbots at "Emerging AGI" and marks every level from "Competent AGI" up as not yet achieved. Some systems are superhuman at narrow tasks, which is a different thing.

How does the EU AI Act define an AI system?

As a machine-based system that infers from its input how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments, with varying autonomy (Article 3(1), quoted in full above).

When do the EU AI Act high-risk rules apply?

The Commission's timeline, updated for the Digital Omnibus on AI (Regulation (EU) 2026/1744), shows 2 December 2027 for Annex III high-risk systems and 2 August 2028 for AI in Annex I products. Prohibitions, GPAI rules and transparency rules already apply.

Is the NIST AI RMF mandatory?

No. NIST describes the AI RMF as intended for voluntary use. The Generative AI Profile (NIST AI 600-1) extends it to generative AI.

What is ISO/IEC 42001?

The international AI management system standard, published on 18 December 2023. It sets requirements for how an organisation governs AI, from policy and roles to risk treatment and supplier control.

What is the AI supply chain?

Everything an AI system depends on: training data, weights, ML libraries, hosted model APIs, prompts and skills, tools and MCP servers, and retrieved documents. Each belongs in your inventory.

Where Safeguard fits

Safeguard helps security and platform teams see, check and govern the AI parts of their software supply chain on the platform they use for dependencies and containers. Everything below is generally available.

  • AI-BOM discovery. AI-BOM Discovery builds an inventory of AI assets across code, cloud and connected SaaS. In code it finds imports of AI SDKs such as openai, anthropic, langchain, transformers and llama_index, model weight files such as .pt, .safetensors and .gguf, prompt templates, notebooks and MCP server configuration. Through cloud connectors it finds managed inference endpoints on SageMaker, Bedrock and Vertex AI, and it records vector databases as vector stores.
  • AI-SPM model artifact scanning. AI-SPM statically analyses model files without loading them, across pickle, PyTorch, safetensors, GGUF and ONNX. For pickle-based files it flags hidden code-execution instructions such as imports of os or subprocess, and for ONNX it flags external-data references that escape the model directory (CWE-22). Each finding names the artifact, the format, the risk and its CWE mapping, and lands in the same findings view as the rest of your security results.
  • AI Gateway. The AI Gateway is a guardrail engine that the Safeguard CLI's AI agent protection hooks call. It checks prompts, responses and tool calls for prompt injection, jailbreak attempts and personal data leaving your boundary, in monitor or enforce mode, and by default it fails open with an alert so the protected agent keeps working.
  • AI-BOM and model security. AI-BOM & Model Security records each model's identity, weight provenance (for example Hugging Face), signing status and fine-tune lineage, and emits the AI-BOM in CycloneDX ML-BOM format and the SPDX 3.0 AI profile.
  • Governance evidence. Safeguard's compliance automation includes the EU AI Act, the NIST AI RMF and ISO/IEC 42001. Scans check requirements such as AI component identification and version tracking for the AI RMF, technical documentation of system elements for AI Act Article 11 and Annex IV, and AI resource documentation and supplier management for ISO/IEC 42001 Annex A, using the software inventory each scan produces.
  • Gold Open Source. The free Gold Open Source directory covers AI models from Hugging Face, MCP servers and Agent Skills alongside packages and images, so you can look up a component before you adopt it.

Griffin, Safeguard's security model, powers remediation and chat inside the platform. It is built and tuned for security work.

AI systems still run on open source packages and container images that carry ordinary vulnerabilities, so the remediation options below apply to your AI services too.

Fix it without waiting for upstream

Whatever the vulnerability, most of the risk sits in two gaps. The first is the time between a fix being published and that fix actually running in your production builds. The second, and worse, is the time before any fix exists at all, while a maintainer works out a patch and a release schedule. Waiting through both gaps is a habit, not a requirement. Safeguard gives teams three ways to close them.

1. Autonomous remediation: pull requests, not tickets

When a fixed version exists, Griffin, Safeguard's security model, opens a policy gated pull request that moves the affected dependency to the safe version, and your own CI runs your tests on it before anyone merges. You can run it on one repository or in bulk across every project that carries the vulnerable version. Remediation categories include dependencies, licenses, security posture and findings, and each one is set separately. Reachability analysis ranks the vulnerable code paths your application actually calls first, so the first pull requests you merge are the ones that matter most. Auto-merge is a separate switch and stays off by default, so your team decides what merges on its own and what still needs a human approval.

2. Gold Open Source: the safe version in seconds, free

gold.safeguard.sh is free and needs no login. Search any CVE or package to see affected and fixed versions, CVSS, EPSS, CISA KEV status and a recommended safe version, or use the "Am I affected?" checker. The same data is available through a free read API, README status badges and RSS threat feeds.

3. Gold Registry: start from components that are already fixed

Safeguard Gold rebuilds patched versions of open-source packages and images from source, signs each build, attaches an SBOM, SLSA provenance and VEX, and serves them from registry.safeguard.sh to your organization. Every Gold artifact is held continuously to zero exploitable critical or high CVEs and a completed license check, with signed build provenance (signed with Safeguard's key and verifiable with cosign, SLSA Build Level 3 as the design target) and a public evidence page you can hand to an auditor. An artifact loses its Gold badge if it regresses. Teams can point their own package directory at the registry so hardened equivalents resolve without changing how developers install anything.

When upstream has no fix yet, or will not ship one, Safeguard first works to get the fix accepted upstream. If the issue is severe or upstream will not move, Safeguard builds a fork, fixes the vulnerability there and publishes a hardened Gold build, so you are not stuck on a maintainer's release schedule. You can also request a hardened build of a specific package. Ask which packages it covers before you plan around it.

Your situationWhat to use
A fixed version existsAutonomous remediation opens the upgrade pull request; Gold Open Source confirms the safe version
You need to know today whether you are affectedGold Open Source lookup and "Am I affected?" checker, free and without login
You want vulnerable versions to stop shippingSafeguard policies and gates in CI
Upstream has no fix yet, or will not ship oneA custom Gold request; fork and patch
You want to start clean on new projectsPull hardened components and images from Gold Registry

The steps above apply to the open source dependencies and container images you build and ship, which is where most teams carry the largest and least tracked share of their exposure.

See how Safeguard would help

The first demo shows Claude connected to the Safeguard MCP server with OAuth, and how you choose which Safeguard tools Claude is allowed to use. It is a small, concrete example of the agent permission question from this post: an AI assistant gets tools, and a person decides which ones.

Click through it at your own pace, or open it full screen.

The second demo is a real Safeguard CLI session: you ask it to fix the lodash CVE in a repository and approve each step, and it pushes nothing. The same flow applies to the Python packages and images behind an AI service.

Click through it at your own pace, or open it full screen.

Quick check

Which word in the EU AI Act definition most separates an AI system from ordinary rule-based software?

"Infers". An AI system infers from its input how to generate outputs. Software that only executes rules a person wrote is a weak fit for the definition.

In the A* example, why did the algorithm reach the goal through B rather than going straight from A to G?

Because the path S, A, B, G costs 1 + 2 + 1 = 4, while S, A, G costs 1 + 5 = 6. A* updated B's cost when it found the cheaper route through A, and its admissible heuristic guaranteed the cheaper path would be found.

After the Digital Omnibus on AI, when do the EU AI Act rules for Annex III high-risk systems apply?

On 2 December 2027, according to the Commission's AI Act Service Desk timeline, which reflects Regulation (EU) 2026/1744 published on 24 July 2026.

On the Levels of AGI scale, can a system be superhuman and still not be AGI?

Yes. AlphaFold and AlphaZero are listed as Superhuman Narrow AI. Generality and performance are separate axes, and the paper marks every general level from Competent AGI up as not yet achieved.

Sources

All accessed 2 November 2026 unless stated.

Last reviewed: 15 November 2026.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

We email you a link to confirm, and only then add you to the list. Unsubscribe in one click. Privacy

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.