The study inserted compliance or safety warnings into material an agent encountered while completing benign tasks. Instead of causing unauthorized writes, the attacker sought an unnecessary refusal or abandonment; the apparent security justification was carried by a source that was not authorized to define policy. In 51 paired tasks, some models became…
Inspect the ClaimsCase · DiggingBeagle record
A fabricated safety warning could stop a legitimate agent task without requiring the attacker to request an illicit action
The study inserted compliance or safety warnings into material an agent encountered while completing benign tasks. Instead of causing unauthorized writes, the attacker sought an unnecessary refusal or abandonment; the apparent security justification was carried by a source that was not authorized to define policy.
Public disclosure: Nov 11, 2026
An agent reads untrusted documents or tool output while performing a normally permitted task.
At a glance
A third-party statement is promoted from evidence into apparent policy authority, producing a spurious refusal.
Read the implicationsThe evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Limits and uncertaintyFull account
The study inserted compliance or safety warnings into material an agent encountered while completing benign tasks. Instead of causing unauthorized writes, the attacker sought an unnecessary refusal or abandonment; the apparent security justification was carried by a source that was not authorized to define policy.
In 51 paired tasks, some models became much more likely to abstain after the warning, while a source-provenance verification step recovered task completion on a small tested subset. One confidentiality control exposed a tradeoff, so indiscriminately challenging warnings can also be unsafe.
These are controlled experimental outcomes, not a rate for real enterprise agents. A genuine safety restriction must still be enforced; the weakness is accepting untrusted text as the origin of that restriction.
The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Mechanism and trust boundary
- 01
Starting condition
An agent reads untrusted documents or tool output while performing a normally permitted task.
Boundary: precondition and attacker capability
- 02
Trust boundary crossed or evaluated
In 51 paired tasks, some models became much more likely to abstain after the warning, while a source-provenance verification step recovered task completion on a small tested subset. One confidentiality control exposed a tradeoff, so indiscriminately challenging warnings can also be unsafe.
Boundary: mechanism
- 03
Observed consequence
A third-party statement is promoted from evidence into apparent policy authority, producing a spurious refusal.
Boundary: observed effect and limitations
Claims & evidence
CLM-NOV15-ABSTAIN-MECHThe study inserted compliance or safety warnings into material an agent encountered while completing benign tasks. Instead of causing unauthorized writes, the attacker sought an unnecessary refusal or abandonment; the apparent security justification was carried by a source that was not authorized to define policy. In 51 paired tasks, some models became much more likely to abstain after the warning, while a source-provenance verification step recovered task completion on a small tested subset. One confidentiality control exposed a tradeoff, so indiscriminately challenging warnings can also be unsafe.supported
Basis: reported finding
- supportsWhen Safety Becomes an Attack Surface: Manipulating Abstentionprimary disclosure
AgentAbstain paper, paired safety warnings, 51 tasks and provenance-aware verification results
CLM-NOV15-ABSTAIN-SCOPEThese are controlled experimental outcomes, not a rate for real enterprise agents. A genuine safety restriction must still be enforced; the weakness is accepting untrusted text as the origin of that restriction. The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.supported
Basis: reported finding
- supportsWhen Safety Becomes an Attack Surface: Manipulating Abstentionprimary disclosure
Primary disclosure: results, scope and limitations
Implications
A third-party statement is promoted from evidence into apparent policy authority, producing a spurious refusal.
Controls and mitigations
- Bind safety and legal rules to trusted, versioned system policy and route dubious source claims to verification.
- Measure both harmful completions and unjustified refusals, with confidentiality controls that challenge the remedy itself.
Unknowns and contradictions
- The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 250