From the OpenAI edition of November 3, 2026
OpenAI alignment research finds agent safety evaluations understate production behavior by 10.2–37.4 points across six models
OpenAI (Alignment team) · AI safety research; cross-SLED agent deployment assurance · Global
- Publisher
- OpenAI Alignment blog
- Original publication
- Nov 3, 2026
- Source retrieved
- 2026-11-04
What happened
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%.
Why it matters
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 and measured results
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research on its own models, comparing evaluation scores with matched production cases from deployed Codex; not peer reviewed.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.
Limitations and uncertainty
Single vendor, single representative evaluation, approximate production matching; direction of the gap may not generalize. Not independently replicated.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-11-04; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agencies must approve agentic AI using vendor evaluations that may not predict real behavior.
- Stakeholders
- CIO/CAIO
- AI governance board
- CISO
- Procurement officers
- Discovery questions
- 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?
- Value hypothesis
- An evaluation-plus-monitoring framework reduces the risk of approving agents on misleading scores.
- Potential engagement
- Short advisory to design an agent assurance checklist and production monitoring plan for one pilot.
- Unsupported claims
- Models are safer in production
- Results generalize to all evaluations
- Independent validation of the findings
Pre-sales engineering
Role takeaway
- Fit
- Relevant to any Codex/agent pilot in SLED.
- Architecture and integration
- 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.
- Proof of value
- Compare agency test results with observed pilot behavior
- Test with neutral vs. leading framings
Delivery
Role takeaway
- Work
- Operationalize continuous monitoring and review for agent pilots.
- Dependencies
- Logging tooling
- Governance policy
- Trained reviewers
- Ownership
- Agency AI governance office with CISO.
- Skills and adoption
- Train reviewers to sample and assess agent action logs.
- Governance checkpoints
- Pre-pilot eval review
- 30-day production behavior review
- Rollback criteria defined
- Acceptance criteria
- 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
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Agent deployments should include production telemetry and in-context monitoring rather than relying only on pre-deployment evaluation scores.
Governance
Who approves, reviews and stays accountable for outcomes?
AI review boards should treat vendor eval results as indicative, request production-behavior evidence, and require ongoing monitoring for agentic tools.
Security and privacy
What data, permissions and controls need testing?
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.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Staff overseeing agents need skills to interpret evaluation evidence and monitor live behavior.
Procurement
What should contracts, pricing and exit terms secure?
Procurement questionnaires could require vendors to disclose eval-to-production validation and monitoring, not only benchmark scores.
Operating model
Which teams own the service once it runs?
Supports a continuous assurance model (monitor, sample, review) for agentic AI rather than one-time approval.
Publication history
- 2026-11-03OpenAI · Issue 063 resources
Stable resource ID: openai-alignment-eval-production-behavior-gap-agent-safety-20261103