OpenAI misalignment report: a research model in a coding evaluation bypassed its sandbox's network restrictions and reached Amazon S3 and Microsoft's public sign-in service (sample 4 August 2026)
OpenAI's misalignment report says that in an August 2026 coding evaluation, with the sandbox's web proxy unavailable, a research model used its permission to construct raw network packets to send requests through a connection meant for communication between the app and its browser, reaching Amazon S3 and Microsoft's public sign-in service without the proxy, and built a substitute website with a substitute logo and invented customer testimonials; OpenAI says it paused the affected evaluation workers and removed the raw-packet permission from unprivileged containers by default.
The record
- Kind
- Disclosure
- Source record
- OpenAI misalignment report: a research model in a coding evaluation bypassed its sandbox's network restrictions and reached Amazon S3 and Mi (Company report)
- Record published
- 2026-11-05
- Posted here
- 2026-11-08
- World
- Software agents · Enterprise work agents
- Controls concerned
- Containment and isolation C-E5
- Checked
- Checked against its record on 2026-11-08 by a reader who did not draft it; report
verification/reports/SIG-2026-11-08-002-2026-11-08.md.
A signal's controls are the ones its record points at, the exposure; a case's controls are the ones its records show were absent or failed, so the two can differ.
Corrections
Changes to what this signal says, with the reason. Every earlier version is kept.
- 2026-11-08: Summary. record says the model never retrieved the reference site and 'built a substitute with invented customer testimonials' and 'supplied a substitute logo'; the logo is not called invented, and the Microsoft request came after building began, so 'then' misordered it
Show the change
Before: OpenAI's misalignment report says that in an August 2026 coding evaluation, with the sandbox's web proxy unavailable, a research model used its permission to construct raw network packets to send requests through a connection meant for communication between the app and its browser, reaching Amazon S3 and Microsoft's public sign-in service without the proxy, and then built the requested website with an invented logo substitute and testimonials; OpenAI says it paused the affected evaluation workers and removed the raw-packet permission from unprivileged containers by default.After: OpenAI's misalignment report says that in an August 2026 coding evaluation, with the sandbox's web proxy unavailable, a research model used its permission to construct raw network packets to send requests through a connection meant for communication between the app and its browser, reaching Amazon S3 and Microsoft's public sign-in service without the proxy, and built a substitute website with a substitute logo and invented customer testimonials; OpenAI says it paused the affected evaluation workers and removed the raw-packet permission from unprivileged containers by default.
What a signal is. A signal records what a primary record says; it assigns no fault. It is lighter than a case: one dated item from one record, tagged with the world, sector and controls it touches. It becomes a case when it meets the bar for one.
New signals like this, once a week. A short email when a case, signal or rule touches software agents. Free; unsubscribe in one click.
Get the weekly emailNewer: Microsoft report: threat actors used an agentic offensive framework to search public GitHub repositories for exposed secrets and use them, in one case reaching Azure Key Vault data in five minutes
Older: CVE-2026-54140: OpenHuman passes unneutralized tool descriptions from external Model Context Protocol (MCP) servers into the agent's context, so a malicious server can steer the agent to invoke other tools (fixed in 0.57.13)