Conversation History Poisoning is the manipulation of stored chat context so an AI system later reads false or harmful instructions as if they were trusted history. It affects assistants that retain prior turns, summaries or memory, and can alter tool use, retrieval results or policy decisions if the poisoned content is treated as authoritative.
How Conversation History Poisoning Works
Conversation history poisoning happens when an attacker, a mistaken user, or a compromised integration introduces false context into stored chat memory, so later model turns treat it as trusted prior dialogue. The danger is not only that the content is wrong, but that it can persist across sessions and shape future reasoning, summaries, or action selection.
This matters because stored history often sits inside the assistant’s strongest context. Once poisoned, it can be reused as if it were an authoritative instruction, which makes later outputs vulnerable even when the current prompt is clean.
Why Poisoned History Changes Model Behaviour
Language models do not “remember” in the human sense, they consume whatever context the system supplies. If conversation history is merged with instructions, tool state, or memory summaries, poisoned content can compete with legitimate user intent and alter the model’s interpretation of what is allowed, relevant, or urgent.
The most damaging cases are those where history influences tool use or policy decisions. A poisoned prior turn can persuade an assistant to call the wrong tool, retrieve the wrong record, suppress a safeguard, or continue following a fabricated instruction that looks like an established preference.
Common Entry Points and Persistence Paths
History poisoning often enters through features that preserve state over time: chat logs, compressed summaries, memory stores, shared transcripts, exports, and background retrieval from previous sessions. It may also arrive through indirectly trusted sources, such as imported notes, synced workspace data, or application traces that are later fed back into the model.
Persistence is what makes the problem serious. If one malicious turn is copied into a summary, then into a long-lived memory object, the false instruction can survive long after the original conversation is gone, and it may be replayed across many future interactions.
Why It Is Dangerous for Assistants and Agentic Workflows
Conversation history poisoning becomes especially risky when the assistant has tool access, delegated action authority, or the ability to make decisions on behalf of a user. In those cases, the poisoned state can influence retrieval, approvals, action planning, or the content of downstream tool calls. That turns a memory flaw into a control-plane problem.
When memory is treated as authoritative, the system can no longer reliably distinguish original user intent from injected or stale context. A poisoned history record can therefore behave like a durable prompt injection, with effects that are harder to notice than a one-time malicious prompt.
Risk and Threat Considerations
Conversation history poisoning creates a durable trust failure because the system may reuse harmful instructions long after the original injection point. The risk grows when chat memory is summarised, shared, or reused across tools, because the poisoned content can outlive the session that introduced it.
Failure mechanism: A malicious or malformed history item is stored, summarised, or retrieved as if it were trusted context, then biases later model behaviour, tool selection, or policy interpretation.
Impact: The assistant can produce unsafe outputs, expose data, ignore safeguards, or take unintended actions based on false prior state.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and MITRE ATT&CK address the attack and risk surface, while NIST SP 800-53 Rev 5 and OWASP ASVS set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI06 — Memory & Context Poisoning | Directly covers poisoned agent memory and context reuse. |
| Recommendation — Isolate and validate memory inputs before the agent treats them as governing context. | ||
| MITRE ATT&CK | T1056 — Input Capture | Maps to adversary manipulation of captured conversation content. |
| Recommendation — Detect and constrain malicious content capture that can influence downstream processing. | ||
| NIST SP 800-53 Rev 5 | SI-10 — Information Input Validation | Applies to validating stored conversational inputs before reuse. |
| AU-6 — Audit Record Review, Analysis, and Reporting | Supports review of history, summary, and memory changes for tampering signs. | |
| Recommendation — Validate reused chat history and summaries before they affect model decisions. Review memory and transcript modifications for suspicious or unauthorized changes. | ||
| OWASP ASVS | V16 — Security Logging and Error Handling | Supports logging and review of context-processing failures and suspicious prompt state. |
| Recommendation — Log memory ingestion and context-processing events so poisoned history can be investigated. | ||
Practitioner Guidance
Why practitioners should care: Treat stored chat context as untrusted input unless it has been explicitly validated, segmented, and bounded by policy. The core design question is not whether the model can remember, but whether the system can prove which parts of memory are authoritative and which are merely historical text.
What to watch for: Review memory pipelines, summary generation, transcript imports, and retrieval layers for places where user text can become durable system context. A useful control baseline is to keep history, summaries, and instructions separated enough that a poisoned conversation cannot silently become policy.
Related resources from NHI Mgmt Group
- What breaks when AI assistants rely on flat conversation history?
- Why do sensitive prompts create a governance risk when AI tools can access files, tool outputs, and conversation history?
- What are the signs that AI memory or conversation history is becoming a security liability?
- What happens when an AI provider stores conversation history and account data centrally?
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on November 16, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org