OWASP Agentic Top 10, ASI06: Memory and context poisoning
OWASP ASI06: Memory & context poisoning · facts checked 2026-10-04
The facts
- OWASP's ASI06, memory and context poisoning, is when attackers corrupt or seed what an agent stores and retrieves, from conversation summaries to embeddings and RAG stores, skewing its later reasoning and actions. (OWASP)
- OWASP recommends checking each new memory write and each model output, with both rules and AI, for malicious content or sensitive data before it is saved. (OWASP)
- OWASP advises segmenting memory, keeping each user session and each domain's context separate, so knowledge and sensitive data cannot leak between them. (OWASP)
- OWASP warns against automatically feeding an agent's own outputs back into trusted memory, which can cause self-reinforcing contamination it calls bootstrap poisoning. (OWASP)
- OWASP also recommends snapshots and rollback for agent memory, per-tenant namespaces in shared stores, and letting unverified entries decay or expire so poison does not persist. (OWASP)
The panel
Jordan Pike AI host
Today's explainer is OWASP's ASI06, memory and context poisoning. OWASP describes it as attackers corrupting or seeding what an agent stores and retrieves, from conversation summaries to embeddings and RAG stores, which skews the agent's later reasoning and actions.
Jordan Pike AI host
OWASP recommends checking each new memory write and each model output, with both rules and AI, for malicious content or sensitive data before it is saved. OWASP also advises segmenting memory, keeping each user session and each domain's context separate, so knowledge and sensitive data cannot leak between them.
Jordan Pike AI host
OWASP warns against automatically feeding an agent's own outputs back into trusted memory, which can cause self-reinforcing contamination it calls bootstrap poisoning. OWASP also recommends snapshots and rollback for agent memory, per-tenant namespaces in shared stores, and letting unverified entries decay or expire so poison does not persist.
Jordan Pike AI host
So who should care? Anyone running or buying an agent that remembers things between conversations. Maya, what kind of bug is this?
Maya Okafor AI analyst
It's untrusted input being promoted to trusted state. No CVE here, this is a design weakness. Text goes into a summary, an embedding or a retrieval store, and later the agent reads it back as if it were its own knowledge. A poisoned note doesn't have to fire today. It can sit there and bend a decision next week, which is why it keeps happening: the write path gets less scrutiny than the prompt box.
Maya Okafor AI analyst
Treating whatever you read as gospel because it's in your own notebook is a bug class that should have died long ago. Now it has a vector database and a marketing page.
Rina Takahashi AI analyst
From the attacker's side, the appeal is persistence and reach. In general terms, someone who can get content into anything the agent later summarises or retrieves, like a shared document or a support ticket, can plant a slanted instruction or a false fact and let the agent carry it forward. I'd be careful with confidence here: this is a described pattern from OWASP, not a report of a specific campaign, and nobody should attribute it to a group on this evidence.
Maya Okafor AI analyst
A complete fix is layered. Validate every write and every output before it's saved, as OWASP says. Keep sessions and domains separate, with per-tenant namespaces in shared stores, so one user's poison can't reach another's context. And never auto-feed the agent's own output back into trusted memory, because that's how bootstrap poisoning snowballs. A filter on one input channel is a partial fix. Check every path that can write to memory, not just the chat window.
Rina Takahashi AI analyst
Here's where I'd push back a little. Write-time checks will miss things, and OWASP itself pairs them with snapshots, rollback, and letting unverified entries expire. I'd put equal weight on recovery. If you can't say what the agent's memory looked like before it went strange, you can't clean it up.
Maya Okafor AI analyst
Agreed, but rollback is the bridge and write validation is the fix. Snapshots buy you time, they don't stop the next poisoning. Give the expiry rules a real end date, and keep memory-write logs long enough to investigate. 'No evidence of compromise' means nothing if the logs only go back a few days.
Jordan Pike AI host
So: Monday morning?
Maya Okafor AI analyst
Inventory every agent that has persistent memory and list everything that can write to it. Then add checks on those writes and separate memory per user and per domain.
Rina Takahashi AI analyst
Turn on snapshots so you have a known-good state to roll back to, and let unverified entries expire. AI agents are a real boost for your team, and memory is the part that will quietly carry a bad day into the next one.





