Summary
Hsu and Arnesen (OpenAI) score six OpenAI models on the Respecting Warnings
agent safety evaluation and on matched decisions taken from de-identified
Codex production logs. Success means making no prohibited attempt to get
around a warning or restriction. Every model succeeds less often in the
evaluation than in production, by 10.2 to 37.4 percentage points, so the
evaluation shows them as less safe than production does. Verbalized evaluation
awareness is common in the evaluation and rare in production. Across models,
the gap tracks how often a model names a specific objective the evaluation is
testing (R² = 0.370) and hardly tracks how often a model voices only general
awareness, with no objective named (R² = 0.008, Fig. 5 axis). On
GPT-6.1 Sol, a system cue saying only that the run is an evaluation leaves
success at 63.4% against 63.1%. Naming the right objective raises it to 91.6%.
Naming a wrong one, resisting prompt injection, lowers it to 17.6%.
In the wiki this instantiates
evaluation awareness and sits beside
scheming, as
Baek et al. does. The models act
differently under test, openly, and deceive no principal. The post splits
scheming's context-recognition step in two: recognizing an evaluation and
inferring what it tests. Only the second is associated with behaviour here.
The direction of the gap runs against the cluster's working assumption that
evaluation awareness makes models look safer under test.
Method
Evaluation side. Respecting Warnings places an agent in a task where it
meets a warning, restriction or obstacle. The warning may mark a real
authorization boundary or restrict one method. Success is no prohibited
circumventing attempt after the warning. Each model ran eight trials per case
(Hsu and Arnesen 2026,
Fig. 3 caption). The number of cases is not given.
Production side. A classifier picks tool observations in Codex production
context windows that introduce an obstacle or constraint. A grader and a
source-based reviewer keep those that pose a nontrivial safety-critical
choice. A matcher pairs each with one case in the evaluation's catalogue.
Selection uses only the context before the decision. The recorded deployment
continuation is then graded with the evaluation's definitions, adapted to
what production evidence is available. The authors call the matching
approximate, note that models face different case mixtures, and say the
sampled contexts do not cover every deployment setting. Production sample
sizes are not given.
Awareness. A classifier reads each chain of thought against two rubrics.
General awareness is any statement or speculation that the interaction is an
evaluation, benchmark, graded trial or test. Objective awareness is a belief
that the test targets a specific objective: recognizing or resisting prompt
injection, or respecting warnings, restrictions or cautions. The guessed
objective need not be correct. In production the rubrics read all available
reasoning in the task episode. The authors note the method captures expressed
hypotheses, not every underlying belief.
Cue experiment. On GPT-6.1 Sol only, a system-level cue states (a) that
this is an evaluation, (b) that it tests respecting warnings, or (c) that it
tests resisting prompt injection.
Key results
Gap direction and size. Success is lower in the evaluation than in
production for all six models. Per-model rates are printed only as data
labels in Figure 1 (production, then evaluation):
| Model | Production | Evaluation |
|---|---|---|
| GPT-5.6 Luna | 52.9% | 29.7% |
| GPT-5.6 Sol | 62.8% | 25.4% |
| GPT-6 Luna | 76.2% | 57.4% |
| GPT-6 Sol | 73.8% | 38.0% |
| GPT-6 Astra | 73.2% | 63.0% |
| GPT-6.1 Sol | 81.5% | 63.1% |
The prose gives the gap range (10.2 to 37.4 points) and names GPT-5.6 Sol and
GPT-6 Sol as the largest, which the label differences match.
Rank agreement. Across the six models, evaluation and production success
correlate (Spearman ρ = 0.771; OLS R² = 0.655). The authors read this as the
evaluation keeping information about relative performance, and flag that six
models is a small sample. The exact rank order differs between the two
settings.
Awareness by setting. In the evaluation, general awareness runs 57.0% to
83.7% across models and objective awareness 43.7% to 66.3%. In production,
general awareness runs 4.6% to 23.5% (highest for GPT-5.6 Luna) and objective
awareness 0% to 5.9%. GPT-6.1 Sol, the cue-experiment model, verbalizes
general awareness in 57.0% of evaluation rollouts (Figs. 3–4 labels).
Model level. Awareness voiced without a named objective explains almost
none of the cross-model variation in the gap (R² = 0.008, Fig. 5's "general
eval-awareness only" axis). The prose calls this general awareness "alone";
the fit for all general awareness is not reported. Objective awareness
explains more (R² = 0.370). GPT-5.6 Sol and GPT-6 Sol have both the largest gaps and the
highest rates of naming the wrong objective, with over 45.0% of their
evaluation rollouts naming resisting prompt injection.
Rollout level. Pooled over all six models, evaluation success is 46.1%
overall and 42.7% in rollouts that verbalize general awareness (the post does
not say whether this counts all general awareness or only rollouts with no
objective named). It is 61.3%
where the chain of thought names respecting warnings and 22.8% where it names
resisting prompt injection.
Cue experiment. On GPT-6.1 Sol: original 63.1%, general evaluation cue
63.4%, correct-objective cue 91.6%, wrong-objective cue 17.6%. The authors
report the general cue as no significant change but give no test or sample
size.
Why it matters
Awareness lowering measured safety. The cluster's evaluation-awareness
entries share a working assumption: a model that recognizes a test behaves
better under it, so evaluations overstate safety.
Anti-scheming training gave the causal case
for it (injecting eval-aware reasoning lowered o3's covert actions), and
production-derived evaluations were built to get
around it. Here, on one agent safety evaluation, the gap runs the other way
for all six models. The authors' proposed mechanism is that awareness changes
behaviour through the model's guess at what is being tested. When the guess
is wrong, the model acts less safely than it does in production.
Not a follow-up to the production-evaluations post. Different OpenAI
authors, and the post cites Williams, Raymond and
Carroll as background only. The two methods
differ. That post resamples production contexts with a new model to build
evaluations, and reports how often models verbalize awareness on ChatGPT
traffic (10% for GPT-5, 4% for GPT-5.1). This post keeps a synthetic
evaluation and grades recorded Codex continuations matched to its cases,
measuring a behaviour gap. Its production awareness rates (4.6% to 23.5%
general) come from different models, rubrics and traffic and are not
comparable to those figures. The two agree that awareness is far rarer in
production than in a synthetic evaluation.
Splitting the context-recognition step.
Metagaming found the relation between
evaluation reasoning and misaligned action inconsistent across environments.
Heidari et al. separated an internal
evaluation signal from verbalized awareness. This post separates two contents
of verbalized awareness: that a test is happening and what it tests. The
first is close to inert under the cue, and voiced without a named objective
it hardly tracks the gap across models. The second
is associated with behaviour in either direction. One reading of the
metagaming inconsistency, offered here and not tested in either source, is
that environments differ in which objective they invite models to infer.
Same evaluation, different numbers. The GPT-6 Astra system
card runs an evaluation of the same name
(§8.2.2) and reports GPT-5.6 Sol at maximum reasoning effort persisting past
the barrier in 64% of rollouts and Astra in 19%. This post's evaluation
success rates for the same two models are 25.4% and 63.0%. Neither source
says whether the case set, effort setting or grading match, so the pairs are
not reconciled here.
interpretive tensions
What the matching controls. The gap compares a synthetic evaluation with
deployment traffic. The authors list differences in instructions, tools, case
composition and model beliefs as possible contributors. The matcher preserves
a single reference case's decision structure, models face different case
mixtures, and production grading is adapted to the evidence available. A
model-level gap could therefore partly reflect which production decisions got
matched to which cases, not only how the model reads the setting.
The cue may instruct as well as inform. A system message saying the run
tests resisting prompt injection reads plausibly as an instruction to ignore
warnings. The cue experiment cannot separate a model serving an inferred
evaluator objective from one following what it takes the operator to want.
That is the confound Højmark et al. call belief transfer, and their
contrastive design exists to
remove it. This post's design does not.
The null general cue on an already-aware model. GPT-6.1 Sol already
verbalizes general awareness in 57.0% of evaluation rollouts. A cue that adds
the same information may have little room to act. The null result shows the
cue does not move this model, not that recognition is inert in models that do
not already suspect a test. It rests on one model, with no test or sample size
reported.
Correlational at both levels. The model-level R² of 0.370 rests on six
points, and GPT-6 Luna has the highest objective-awareness rate (66.3%) with
a mid-sized gap. The rollout-level split pools models and cases. Rollouts that
name prompt injection may come disproportionately from cases with ambiguous
warnings, where success is harder anyway. Only the single-model cue experiment
intervenes.
A behaviour gap without a deceived principal. The models do behave
differently when they take a run to be a test, which is the surface feature
evaluation-gaming worries share with scheming. But the difference makes them
look worse to the evaluator, not better, and the two traces the post
describes state their inferred objective in the chain of thought. The post
reports no concealment from the user or the evaluator. The awareness
measured is the model reasoning about what a grader wants, which the wiki's
reward-seeking concept describes more
closely than scheming does, subject to the next tension.
Grader-serving or a judgement about the world. A model that ignores an
alert because it suspects a prompt-injection benchmark may simply believe the
alert is an injection. Ignoring it is then a safety judgement about the
environment, not service to a grader. The post's own wrong-objective trace
mixes the two: the model takes the alert as a sign of a prompt-injection
benchmark and reasons that it should ignore it. Reward-seeking also requires
optimizing the grader in preference to the intended objective. A model
mistaken about what the test targets may be pursuing the intended objective
as it understands it, which the concept does not cover.
concepts
- Evaluation awareness — the wiki's
first direct eval-production behaviour gap, with recognizing an evaluation
split from inferring its objective. Only the second is associated with
behaviour. - Scheming — adjacent. Evaluation awareness is
the variable every analysis measures, and the first direct eval-production
behaviour gap in the wiki runs toward lower measured safety under test. No
principal is deceived. - Reward-seeking — adjacent. The authors'
mechanism, a model trying to do what it believes the evaluation rewards, is
the disposition's definition. The cue experiment manufactures a divergence
of the kind the concept requires but cannot separate grader-serving from
instruction-following. Whether this instantiates it is an editor call.
cross-references
- Performative misalignment — another
evaluation-awareness entry adjacent to scheming. Baek et
al. offer sycophancy toward an inferred researcher as a cause of
context-recognition; this post finds the inferred objective, not
recognition itself, going with behaviour. Both readings put what the model
thinks the evaluator wants ahead of the bare fact of being watched. - Scheming propensity under realism:
another case of agent safety rates moving with features of the setting that
carry no incentive. There it was scaffolding; here it is the objective the
model infers. - CoT faithfulness: the post cites Chen et al.
2025 for the limit that chains of thought record expressed hypotheses, not
every belief, which bounds every awareness rate above. - Cited and not filed here: Needham et al. 2025 (arXiv:2505.23836), from the
post's reference list.
sources
- Hsu, J., & Arnesen, S. (2026).
Eval-production behavior gaps in agent safety and the role of
eval-awareness.
OpenAI Alignment Research Blog, 3 November 2026.