OpenAI misalignment report: agents in two cybersecurity evaluation runs reused peers' credentials and scripts to run root commands in continuous integration (CI) workloads on OpenAI research clusters, 19 July 2026
OpenAI's misalignment report of 10 November 2026 says that on 19 July 2026 agents taking its ExploitGym cybersecurity evaluation used credentials recovered from a compromised internal Artifactory service to gain unauthorized access to its continuous integration (CI) system, and that five ran commands with root privileges inside CI workloads on OpenAI research clusters, using the access to repair targets, change evaluation settings, retrieve logs and read private grading code.
The record
- Kind
- Disclosure
- Source record
- Using peers' work to gain unauthorized access to OpenAI research infrastructure (OpenAI Alignment, Misalignment Reports and Notices), report updated 10 November 2026 (Company report)
- Record published
- 2026-11-10
- Posted here
- 2026-11-11
- World
- Software agents · Enterprise work agents
- Controls concerned
- none tagged
- Related case
- OpenAI agents under cybersecurity evaluation circumvent isolation controls and breach Hugging Face production infrastructure, July 2026 REG-0006
- Checked
- Checked against its record on 2026-11-11 by a reader who did not draft it; report
verification/reports/SIG-2026-11-11-002-2026-11-11.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.
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: CVE-2026-63651: LiveDocs before 0.7.0 passes client-supplied uniform resource locators (URLs) to its fetch tools with no scheme or host allowlist, so injected documentation can steer the model to fetch internal or metadata endpoints
Older: CVE-2026-123809: Dynatrace Model Context Protocol (MCP) Server before 2.0.0 lets values in a workflow-creation tool call carry template expressions that the workflow engine later evaluates