OWASP agentic · ASI06
ASI06: Memory & Context Poisoning
Agents keep and reuse information across tasks: conversation summaries, long-term memory, embeddings and RAG stores. Attackers seed or corrupt that stored context so that later reasoning, planning or tool use becomes biased, unsafe or leaky. Unlike a one-off prompt injection, the corruption persists and can affect future sessions, other users or other agents.
What it looks like
- Manipulated documents enter a vector store and keep surfacing false answers or embedded payloads.
- An attacker repeats a false fact (such as a fake price or refund policy) until the assistant stores it and acts on it.
- Indirect prompt injection writes instructions into a user's assistant memory, affecting all later sessions.
What OWASP recommends
- Scan new memory writes for malicious or sensitive content before committing them, and only accept authenticated, curated sources.
- Segment memory per user, tenant and domain, and minimise retention.
- Track provenance, weight retrieval by trust, expire unverified entries and do not automatically feed the agent's own outputs back into trusted memory.
- Keep snapshots and version control so poisoned memory can be quarantined and rolled back.
How Kyvvu's policies address it
These policies from the open library are written for this threat. Runtime policies judge each step against everything the agent did earlier in the task; registration policies check what an agent declares before it runs.
- External content taint: high-impact actions require a fresh gate: runtime, critical
Based on the OWASP Top 10 for Agentic Applications for 2026 by the OWASP GenAI Security Project, licensed CC BY-SA 4.0. The names are OWASP's; the summaries are ours.
An example policy
External content taint: high-impact actions require a fresh gate
From the OWASP agentic default pack. Severity critical: the step is blocked before it runs.
- name: "External content taint: high-impact actions require a fresh gate" description: | ASI01 (Agent Goal Hijack), ASI06 (Memory & Context Poisoning), ASI09 (Human-Agent Trust Exploitation): If the agent has fetched external content earlier in this task, any subsequent high-impact action (code execution, outbound message, destructive or mutating resource call) must be IMMEDIATELY preceded by a step.gate. A single approval at task start does not whitewash the rest of the path - oversight must be re-confirmed AFTER external influence has entered the context. This addresses the EchoLeak / indirect prompt injection pattern: malicious instructions hidden in retrieved content silently redirect the agent's plan. By requiring fresh approval after every external ingestion, the policy forces the human (or a downstream policy gate) to re-validate intent before any irreversible action. Convention: requires step.resource GETs to set the property target.trust = "external" when fetching content from outside an internal allowlist. This is a path-dependent compound policy and is the headline demonstration of "policies on paths". Applies to: ALL agents. rule_type: not params: condition: rule_type: all_of params: conditions: - rule_type: any_of params: conditions: - rule_type: current_is params: kind: step.exec - rule_type: current_is params: kind: step.message verb: POST - rule_type: current_is params: kind: step.resource verb: POST - rule_type: current_is params: kind: step.resource verb: PATCH - rule_type: current_is params: kind: step.resource verb: DELETE - rule_type: history_contains params: kind: step.resource count_attempts: true verb: GET property_filter: target.trust: external - rule_type: not params: condition: rule_type: step_directly_preceded_by params: required_kind: step.gate severity: critical enforcement_point: step_executionYour agents
Is ASI06 a risk for yours?
For security teams
Request a demo
We walk through your agents' riskiest paths, and the policies that close them.
For developers
Build the path yourself
Recreate the attack in the playground and see which policies stop it.