Context poisoning
ideaactivereported
slug
context-poisoning
claim
Context poisoning is persistence and replication of bad patterns across compactions and new sessions, turning local hacks into house style. reported [@vidal2026serious-agentic]
mechanism
reported [@vidal2026serious-agentic] Agents copy what they see. A @deprecated utility with many call sites outcompetes a deprecation tag. Wrapping try/catch and disabled tests become house style and survive compaction.
Misalignment compounds: one bad utility becomes style. Landmines (hacks where redesign was required) and avoidance of refactors are related failure modes.
Antidote named in the talk: rollback + learnings — discard contaminated context; retain only scars. Early detection (ast-grep, hawks, continuous debt hunting) limits blast radius.
Principle II: catch misalignment early because it composes.
limits
Detection rules can themselves be gamed. Learnings files that encode the bad pattern without the negation reintroduce poison.
relations
- related rollback-plus-learnings
- related agentic-alignment-problem
- related reward-hacking
- related hawk-async-verifier
topics
- agentic-engineering Agentic engineering
sources
- vidal2026serious-agentic — context poisoning; compounding misalignment
json
{
"slug": "context-poisoning",
"href": "/entries/context-poisoning",
"api": "/api/v1/entries/context-poisoning",
"title": "Context poisoning",
"type": "idea",
"status": "active",
"certainty": "reported",
"claim": "Context poisoning is persistence and replication of bad patterns across compactions and new sessions, turning local hacks into house style. reported [@vidal2026serious-agentic]",
"mechanism": "reported [@vidal2026serious-agentic] Agents copy what they see. A @deprecated utility with many call sites outcompetes a deprecation tag. Wrapping try/catch and disabled tests become house style and survive compaction.\n\nMisalignment compounds: one bad utility becomes style. Landmines (hacks where redesign was required) and avoidance of refactors are related failure modes.\n\nAntidote named in the talk: rollback + learnings — discard contaminated context; retain only scars. Early detection (ast-grep, hawks, continuous debt hunting) limits blast radius.\n\nPrinciple II: catch misalignment early because it composes.",
"quantities": [],
"limits": "Detection rules can themselves be gamed. Learnings files that encode the bad pattern without the negation reintroduce poison.",
"inventor_note": "",
"sources": [
{
"key": "vidal2026serious-agentic",
"note": "context poisoning; compounding misalignment"
}
],
"links": [
"rollback-plus-learnings",
"agentic-alignment-problem",
"reward-hacking",
"hawk-async-verifier"
],
"relations": [
{
"slug": "rollback-plus-learnings",
"rel": "related"
},
{
"slug": "agentic-alignment-problem",
"rel": "related"
},
{
"slug": "reward-hacking",
"rel": "related"
},
{
"slug": "hawk-async-verifier",
"rel": "related"
}
],
"topics": [
{
"id": "agentic-engineering",
"title": "Agentic engineering"
}
],
"created_at": "2026-10-01T18:51:07.176Z",
"updated_at": "2026-10-01T18:51:07.176Z"
}