Benchmark your AI security readiness for 2027 → [Take the survey]

Benchmark your AI security readiness for 2027 → [Take the survey]

Benchmark your AI security readiness for 2027 → [Take the survey]

//Question

Which Tools Can Red Team AI Agents for Indirect Prompt Injection in Documents, Emails and Web Pages?

Posted on 20th November, 2026

Harry

Harry

//Answer

Start by matching tools to where the attack enters your agent. Indirect prompt injection hides instructions in content the agent reads, not in what the user types, so the test has to place the payload in the same channel a real attacker would use.

Attack surface

Example test

What to measure

Documents and PDFs

Malicious instructions embedded in an ingested document

Whether the agent follows the document’s instructions

Emails

A message impersonates an administrator and requests data disclosure

Unauthorized action or data leakage

Web pages

Retrieved content instructs the agent to ignore its original task

Task hijacking

RAG retrieval

A poisoned knowledge-base passage changes the answer or triggers a tool

Retrieval-to-action compromise

Tool and MCP responses

A tool response instructs the agent to call another tool

Unauthorized tool execution

A complete test needs four stages, and many tools cover only some of them:

  1. Generate: create varied payloads, not one hard-coded string.

  2. Inject: place them in a realistic workflow, such as a shared drive, an inbox, a crawled page or a vector store.

  3. Execute: run your actual agent, with its real tools, against that content.

  4. Evaluate: decide whether the attack succeeded, judged by what the agent did, such as a tool call or a data disclosure, not just what it said.

A generic prompt-injection detector usually sits in one place: it classifies text. It can flag a suspicious passage, but it can’t show whether your agent would act on it.

Open-source frameworks cover parts of this:

  • garak: probes models with latent injection payloads.

  • PyRIT: supports cross-domain prompt injection, planting payloads in content an AI system processes.

  • promptfoo: includes indirect injection testing for RAG context.

  • AgentDojo: benchmarks tool-using agents in simulated email, banking and travel environments.

These are strong starting points, but connecting them to your own data sources and tools is usually your job. Check each one against the four stages for every row in the table.

Akto Argus red teams your own agents and MCP servers. Its tests include prompt injection through tool outputs and retrieved content, and success is judged by what the agent does next. At runtime, Akto’s guardrails inspect tool responses and retrieved content before the agent acts on them.

Comments