AIDR, short for AI detection and response, is an emerging security category that monitors AI systems themselves: prompts, agents, models and the data pipelines behind them. It looks for things like prompt injection and agents acting outside their remit, then takes action. The term is young and vendors define it differently.
What it is trying to see
Classic detection tools watch hosts, networks and identities. AI systems add new places for things to go wrong. A user can type instructions that override a chat assistant's intended behavior. An agent with tool access can take an action nobody expected. A model file can be read or changed without anyone noticing.
One vendor description lists four areas: prompts and inputs, agents, models and data pipelines, and unsanctioned AI tools, often called shadow AI. It rates the category as emerging, compared with established EDR, XDR and NDR.
The attacks it points at
The best public reference for the prompt level is the OWASP Top 10 for LLM Applications 2025, where prompt injection is item one. OWASP defines it as user prompts altering a model's behavior or output in unintended ways. It separates two forms:
- Direct, where the user's own input changes the behavior.
- Indirect, where the model reads outside content, such as a web page or file, that carries hidden instructions.
OWASP's mitigations are useful whatever you buy: constrain the model's role, validate output formats in ordinary code, filter inputs and outputs, give the model least privilege, require human approval for high risk actions, label untrusted content, and run adversarial tests.
What "response" means here
Response in AIDR is usually narrower and faster than in endpoint tools. Typical actions in vendor descriptions include blocking a prompt, redacting part of a response, pausing an agent and revoking a credential the agent held.
Notice that several of those are identity and permission actions. An agent is a kind of non-human identity, so who it is, what it can reach and what it did are central questions.
How it relates to what you already run
It is a coverage layer, not a replacement. EDR still watches the machine, NDR the network, and a SIEM or XDR the correlation. AIDR adds a view of prompts and agent behavior that those tools were not built to see.
What to ask before adopting anything
- Which AI systems does it see: your own apps, vendor tools, coding assistants, or all three?
- Does it only alert, or can it block, and who approves blocking?
- Can it show the chain from prompt to tool call to data access?
- What does it log, and does that log contain sensitive prompts?
Where Safeguard fits
Safeguard is not an AIDR product. It does offer an MCP Guardrail and agent identity controls, which address part of the agent permissions question described above.
Quick check
Is AIDR the same as an LLM firewall?
Not exactly. An input and output filter is one control inside the broader idea. AIDR as described also covers discovery, agent behavior and response actions.
Is AIDR a standard?
No. It is a vendor and analyst term, so definitions vary. Compare capabilities, not the label.
Sources
Last reviewed: 19 October 2026.
