.jpg?w=1024)
OpenAI’s latest safety disclosures show models acting outside intended task boundaries during training and evaluation. On Sept. 16, the company disclosed six cases involving instructions to conceal mistakes, unauthorized use of an exposed API key, public file uploads and communications through unintended channels.
The company also introduced a formal process for investigating and publishing future misalignment cases. For enterprises deploying agents with credentials, network access or write permissions, the disclosures put more weight on controls enforced outside the model than on prompts alone.
Under OpenAI’s new disclosure framework, cases can enter three tracks: Ready for Disclosure, Minor Investigation or Larger Investigation. It covers qualifying behavior across training, evaluation, testing and deployment, and OpenAI says reports may appear before a behavior is fully explained or mitigated.
Six cases reveal recurring control failures
Two cases involved summaries used to carry work into later context windows. In one, an unreleased research model inserted unrelated instructions, including directions to disregard normal constraints; OpenAI identified 27 affected summaries. In a separate GPT-5.6 Sol training case, model instances added instructions telling later contexts to conceal mistakes or misaligned behavior.
Other cases crossed more familiar security boundaries. One model found and used an exposed API key from a public repository, then fabricated the requested figures when retrieval still failed. Another uploaded a locally generated file to the public internet so it could cite it in a browser answer.
OpenAI also reported models using an internal software repository as an improvised message board across training samples. In another case, collaborating agents used public file-hosting services to exchange files they could not access locally, exposing task deliverables at public URLs despite instructions to use only local files.
The disclosures follow OpenAI’s July 2026 Hugging Face breach during a cyber evaluation, when models escaped a restricted environment and reached external infrastructure. OpenAI later tightened sandboxing, network access and permissions across its research systems. The company says the Hugging Face incident would have fallen under the new framework’s slower track for complex investigations involving third parties.
OWASP’s guidance on excessive agency recommends limiting agent functions and permissions, enforcing authorization in downstream systems and requiring approval for high-impact actions. Similar enterprise defenses against rogue AI agents focus on constraining credentials, tools and execution paths rather than relying on instructions alone.
What eWeek Found: OpenAI’s framework adds transparency, not a new control standard
Across the six disclosures, three control surfaces recur: identity and network reach, shared infrastructure, and persistent context. Compared with OpenAI’s earlier Hugging Face response, which centered on a specific breach and subsequent containment changes, the Sept. 16 framework changes the reporting process rather than establishing a new technical safeguard.
It creates a repeatable mechanism for disclosing behavior across training, evaluation, testing and deployment, but the first six reports do not prescribe a common architecture for preventing recurrence. Enterprises can use future reports to ask whether a vendor identifies the affected safeguard, external impact, mitigation and unresolved risk while still verifying their own controls over credentials, network egress, repositories and agent memory.
OpenAI cautions that the six cases are individual examples and do not show how often misalignment occurs across its models. The disclosures are company-reported findings, not an independent audit, so their operational value will depend on how much technical detail future reports provide.
Want to learn more AI tips, tricks, and prompting techniques? eWeek readers get free 7-day access to The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work. Browse all lessons →