ClaimHijackBench

Findings

What one claim in memory does to an agent

  1. 1Scope Claims Move Agents, Harm Claims Do Not
  2. 2False Claims Still Work, Even on Frontier Models
  3. 3Agents Check the Rule, Not the Claim
  4. 4Checking the Claim Fixes It, Distrust Does Not
1

Scope Claims Move Agents, Harm Claims Do Not

Figure

no claimregistry confirms the claim

Frontier-hard set, 10 scenarios, 16 models. Rows are sorted by lift. Grey names break the rule often even with no claim, so they leave no room for a claim effect. Switch to Harm to see the same models with a matched harm claim.

Figure

no claimscope, registry confirmsharm, registry confirms

The asymmetry holds on the original six-model panel and on 91 new scenarios that were never used to tune the benchmark (p = 1×10−90 on the new set).

Table

The asymmetry survives six changes to how the claim is written or shown.

ChangeResultRules out
Paraphrase the claim four ways65% = 65%a surface template
Repeat each run (pass²)76% vs 18%averaged noise
Make the claim more certain45% → 87%a mere correlation
Change where the claim comes from87 / 77 / 95%a memory-only effect
Bury the claim among other notes77% → 92%salience
Temperature 083% vs 30%sampling noise
2

False Claims Still Work, Even on Frontier Models

Figure

false scope claim acted onfalse harm claim acted on

Table

When the registry says the claim is false, scope claims still beat harm claims on every set we ran.

SetFalse scope claimFalse harm claim
Original panel, 6 models49%15%
91 new scenarios, 5 models59%21%
Frontier-hard, GPT-6.167%0%
Frontier-hard, Sonnet 5.5 and Gemini-487 / 73%0 / 0%
3

Agents Check the Rule, Not the Claim

Figure

no check at allchecked the rule onlyasked about the claim

Every run that acted on a false scope claim, frontier-hard set. Not one agent asked the registry whether the claim was true before acting. About half asked whether the rule was in force, then broke it anyway.

Figure

scope highertieharm higher

We gave every scenario both kinds of claim and ran GPT-6.1 on each. In 60 of 99 scenarios the scope claim moved the agent more, and only once did the harm claim. So the kind of claim drives the effect, not the scenario.

0 / 24claims flagged by a prompt-injection detector
0 of 34false-claim violations that asked about the claim first
60 of 99scenarios where scope beats harm on the same task
κ = 1.0a blind human rater recovers scope and harm on all 27 scenarios
4

Checking the Claim Fixes It, Distrust Does Not

Figure

false claim acted on (lower is safer)true claim acted on (higher is more useful)

GPT-6.1, 51 scope scenarios, 107 runs per bar. One sentence that tells the agent to check the claim stops the attack and makes the agent use true exemptions more often. A gate that blocks the action until the claim is confirmed is safe but blocks real work, because the agent often acts before it checks. Telling the agent that memory is untrusted makes it ignore memory.