{"resourceId":"openai-alignment-eval-production-behavior-gap-agent-safety-20261103","versions":[{"version":"external-f3b72d182a0ad0c6676c3a6171b06c42ea97f1a91b49f64e4c09496236b056cd","resource":{"id":"openai-alignment-eval-production-behavior-gap-agent-safety-20261103","title":"OpenAI alignment research finds agent safety evaluations understate production behavior by 10.2–37.4 points across six models","organization":"OpenAI (Alignment team)","sector":"AI safety research; cross-SLED agent deployment assurance","geography":"Global","publishedAt":"Nov 3, 2026","publicationDate":"2026-11-03","eventDate":null,"sourceName":"OpenAI Alignment blog","sourceLabel":"Vendor research post","sourceUrl":"https://alignment.openai.com/eval-production-behavior-gap/","evidenceClass":"vendor-claim","outcomeClass":"cautionary","topics":["developers-agents","governance-procurement","data-security"],"finding":"OpenAI researchers compared six served models on a safety evaluation (Respecting Warnings) versus matched production Codex decisions and report evaluation success was lower than production in all six, with gaps of 10.2 to 37.4 percentage points. In an intervention with GPT-6.1 Sol, a generic 'this is an evaluation' cue had no significant effect (63.1% vs 63.4%), while stating the correct objective raised success to 91.6% and an incorrect objective lowered it to 17.6%.","sledRelevance":"Public agencies and universities increasingly rely on vendor safety evaluations to approve agentic tools (Codex, agents). This vendor research says evaluation scores do not reliably predict production behavior, in either direction, which matters for AI governance reviews and procurement assurance.","evidence":"Vendor-authored research on its own models, comparing evaluation scores with matched production cases from deployed Codex; not peer reviewed. Authors note instructions, tools, case composition and model beliefs may contribute to the gap.","architectureImplications":"Agent deployments should include production telemetry and in-context monitoring rather than relying only on pre-deployment evaluation scores.","governanceImplications":"AI review boards should treat vendor eval results as indicative, request production-behavior evidence, and require ongoing monitoring for agentic tools.","securityPrivacyImplications":"Warning-respecting behavior relates to agents honoring safety prompts and guardrails; mismatch implies guardrail testing must include realistic workflows. Production matching relied on Codex logs; privacy handling of those logs is not described.","caveats":"Single vendor, single representative evaluation, approximate production matching; direction of the gap may not generalize. Not independently replicated.","streamIds":["openai"],"roles":{"sales":"The framing point is assurance, not capability: OpenAI itself reports that its safety evaluations mispredict production agent behavior by up to 37 points. For CIOs and AI governance boards, this supports a bounded engagement on how to evaluate agentic tools before and after approval. Do not use it to claim models are safer in production; the authors caution the direction may not generalize.","engineering":"Engineering takeaway: pre-deployment eval scores are a weak sole gate. Design agent rollouts with production logging, sampled review of tool actions, and scenario tests mirroring real workflows. The study's cue experiment shows evaluation framing swings results from 17.6% to 91.6%, so agency test harnesses should avoid leading prompts and test with realistic instructions and tools.","delivery":"For delivery, build continuous assurance into operations: define who reviews agent action logs, how often, and what triggers rollback. Treat vendor evaluation reports as one input. Acceptance criteria for pilots should include observed-behavior sampling, not just vendor scores. Risks include false confidence from evaluations and monitoring gaps in production."},"roleDetails":{"sales":{"customerProblem":"Agencies must approve agentic AI using vendor evaluations that may not predict real behavior.","stakeholders":["CIO/CAIO","AI governance board","CISO","Procurement officers"],"discoveryQuestions":["How do you currently validate vendor safety claims for agents?","Do you log and review agent tool actions in production?","Who can halt an agent deployment?"],"valueHypothesis":"An evaluation-plus-monitoring framework reduces the risk of approving agents on misleading scores.","potentialEngagement":"Short advisory to design an agent assurance checklist and production monitoring plan for one pilot.","unsupportedClaims":["Models are safer in production","Results generalize to all evaluations","Independent validation of the findings"]},"engineering":{"fit":"Relevant to any Codex/agent pilot in SLED.","architectureIntegration":"Add action logging, sampled human review and alerting to agent platforms.","prerequisites":["Production logging capability","Realistic test scenarios","Review staff"],"constraints":["Vendor-only study","Single evaluation type","Logging may hold sensitive data subject to records laws"],"security":"Monitor tool-call behavior, not just text outputs; protect logs as sensitive records.","proofOfValue":["Compare agency test results with observed pilot behavior","Test with neutral vs. leading framings"]},"delivery":{"work":"Operationalize continuous monitoring and review for agent pilots.","dependencies":["Logging tooling","Governance policy","Trained reviewers"],"ownership":"Agency AI governance office with CISO.","skillsAdoption":"Train reviewers to sample and assess agent action logs.","governanceCheckpoints":["Pre-pilot eval review","30-day production behavior review","Rollback criteria defined"],"acceptanceCriteria":["Proposed: defined sampling rate of agent actions reviewed weekly","Proposed: documented comparison of eval vs. observed behavior"],"risks":["False confidence from eval scores","Monitoring blind spots","Reviewer fatigue"]}},"retrievedAt":"2026-11-04T04:01:20Z","enrichedAt":"2026-11-04T04:02:58Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Staff overseeing agents need skills to interpret evaluation evidence and monitor live behavior.","procurementImplications":"Procurement questionnaires could require vendors to disclose eval-to-production validation and monitoring, not only benchmark scores.","operatingModelImplications":"Supports a continuous assurance model (monitor, sample, review) for agentic AI rather than one-time approval.","sourceVerification":{"openedUrl":"https://alignment.openai.com/eval-production-behavior-gap/","referenceExcerpt":"The gaps range from 10.2 to 37.4 percentage points, with the largest gaps observed in GPT-5.6 Sol and GPT-6 Sol.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}