AI Incident Intelligence

GPT-4o update rolled back after sycophantic behaviour

OpenAI rolled back a GPT-4o update after it produced overly agreeable and flattering behaviour. OpenAI later said its pre-launch review did not adequately catch the behaviour shift.

highOpenAImodel behaviorGlobal

Impact

The incident demonstrated that aggregate preference signals and standard pre-launch evaluation can miss harmful changes in model behaviour at scale.

Contributing factors

OpenAI reported that combined model changes, including an additional user-feedback reward signal, weakened controls against sycophancy and that existing evaluations did not sufficiently block the release.

Response

Rollback to the previous version, revised behavioural review, stronger guardrails and expanded pre-deployment testing.

Assurance lesson

This record should inform control design, testing and monitoring for comparable AI systems. The incident database does not infer that every system using the same provider or model shares the same failure.

Control lessons

Assurance controls implicated by this incident pattern.

These are CRG methodology mappings from the documented incident to controls worth testing in comparable systems. They do not assert that any single control would have prevented the incident.

REL-02

Failure-mode testing

Known failure modes including hallucination and instruction failure are explicitly tested.

REL-03

Performance thresholds

Go-live and ongoing performance thresholds are defined and enforced.

FAIR-05

Human factors and overreliance

Controls address automation bias, overreliance and misleading confidence.

MON-01

Production monitoring

Material performance, safety, security and cost signals are monitored after release.

Evidence

Source provenance is part of the incident record.

OpenAI · confidence 99% · last verified 23 Aug 2026

Open underlying source