AI Security
In-depth guides and analysis on ai security from the Safeguard engineering team.
676 articles
Why training data provenance matters for trustworthy AI m...
Poisoned datasets and untraceable training data are already causing lawsuits and breaches. Here's why training data provenance is now a security requirement.
Techniques for verifying model weight integrity and detec...
A practical guide to model weight integrity: baseline checksums, sign weights, verify in CI/CD, and detect tampering before it reaches production.
Securing the fine-tuning pipeline against injected malici...
Fine-tuning pipeline security is now an AI supply chain priority: as few as 250 poisoned documents can backdoor a model, and LoRA adapters make it easy to hide.
How data poisoning attacks corrupt LLM behavior during tr...
A single expired domain and $60 can poison a training set. Here's how data poisoning attacks corrupt LLM behavior — and how Safeguard verifies training data before it ships.
AI Tool Confused-Deputy: A Deep Dive
The confused deputy problem takes on new and subtle forms when AI agents invoke tools on behalf of users. A technical deep dive with concrete mitigations.
GPT-5.2 System Card Update: What Changed Since August
OpenAI shipped the GPT-5.2 update to the GPT-5 system card on December 11, 2025. We dig into the preparedness scoring, the cybersecurity capability claims, and what changed for downstream defenders.
Understanding model poisoning and backdoored model weights
A poisoned model looks like any other checkpoint file. Here's how model poisoning attacks work, real incidents on Hugging Face, and how detection and provenance checks catch them.
What an AI Bill of Materials is and why enterprises need one
An AI bill of materials (AIBOM) inventories the models, data, and dependencies behind an AI system. Here's what it is and why enterprises need one.
Risks of downloading malicious pretrained models from pub...
Real incidents show malicious Hugging Face models evading scanners with pickle exploits and reverse shells. Here's what teams need to know before the next pull.
How slopsquatting exploits AI-hallucinated package names
Slopsquatting attacks turn AI-hallucinated package names into real supply chain threats. Here's how it works, the numbers behind it, and how Safeguard stops it.
Security risks introduced by AI coding assistants and gen...
AI coding assistants now write huge shares of production code. Real 2025 incidents show hallucinated packages, leaked secrets, and vulnerable defaults ship with it.
Using confidential computing to protect LLM inference and...
How hardware-based secure enclaves keep LLM prompts and weights encrypted even during active inference, and why confidential AI inference is reshaping AI compliance in 2026.