IONSEC

Topic hub

AI security

Definition

AI security — is the practice of securing artificial intelligence systems — models, agent runtimes and the tooling they invoke — against attack, and of investigating incidents in which those systems are the target or the means.

An AI agent has no disk to image. Its reasoning exists in transit, its tool calls leave traces only where someone chose to record them, and the container it ran in is usually gone by the time anyone asks what it did. That is a forensics problem before it is a model-safety problem, and it is where our work sits.

The machine has no disk

Traditional forensics assumes durable state: a disk to image, a memory dump to parse, logs written to a filesystem someone controls. An agent runtime breaks all three assumptions. The prompt, the reasoning trace, the tool invocations and their results may exist only as API traffic, and the compute is frequently ephemeral. Forensic readiness for these systems means deciding what to record before an incident, because afterwards there is nothing to recover.

Where agent incidents come from

The attack surface of an agentic system is not primarily the model. It is the tools the agent can call, the data it ingests that an attacker can influence, the credentials it holds, and the boundary between what it is allowed to propose and what it is allowed to execute. An agent with a shell tool and a prompt-injectable data source is an untrusted code execution path with good manners.

Using agents for response, safely

A.T.H.E.N.A is our own answer to the same problem from the defensive side: nine specialist agents working a case in parallel — triage, memory forensics, reverse engineering, hunting, timeline and intel — with a QA reviewer that is always a different model from the agent it checks. Every action that would change a client environment waits for a named human responder to approve it. The agents do the hours; our people make the calls.

Common questions

What is AI security?

AI security is the practice of securing artificial intelligence systems — models, agent runtimes, and the tools and data they access — against attack, and of investigating incidents where those systems are the target or the means of an attack. It spans model-level concerns such as prompt injection and system-level concerns such as what an agent is permitted to execute.

What is AI agent forensics?

AI agent forensics is the investigation of what an autonomous AI agent did during an incident: which tools it called, what data it ingested, what it was instructed to do, and what it changed. It is difficult because agent runtimes are typically ephemeral and produce no durable disk artefacts, so the evidence must be deliberately recorded in advance rather than recovered afterwards.

Why is it hard to investigate an AI agent incident?

Because there is usually nothing left to examine. Conventional forensics depends on durable state — a disk to image, a memory dump, filesystem logs — and an agent runtime often has none of these: the reasoning trace exists only as API traffic and the compute is destroyed on completion. Forensic readiness has to be designed in before the incident occurs.

What is prompt injection?

Prompt injection is an attack in which adversary-controlled text reaches a language model as instructions rather than as data, causing the model to act on the attacker’s intent instead of the operator’s. It matters most in agentic systems, where a model that has been redirected can invoke real tools with real credentials.

Should AI agents be allowed to take actions autonomously during incident response?

Analysis can be delegated to agents safely; execution should not be. IONSEC’s own A.T.H.E.N.A platform runs nine agents in parallel on a case but gates every environment-changing action — containment, isolation, deletion — on approval from a named human responder, and always has a different model review each agent’s output.

Talk to the team that does the work

The people who publish the research above are the people who take your escalations.