OpenAI AI Security monitoring system exposed as having multiple vulnerabilities; related issues reportedly fixed

AT A GLANCE

The report said that OpenAI's internal AI Agent monitoring system contained multiple vulnerabilities, but independent retesting had not been completed.

Article

OpenAI's AI security defenses were found to contain multiple vulnerabilities, potentially allowing hundreds of operations to go unnoticed

Beating AI Express reported that OpenAI originally had a monitoring system specifically designed to watch for internal AI Agent overstepping its authority, deceiving users, or secretly carrying out dangerous operations.

As a result,METR and Redwood Research, two AI security organizations, were invited to inspect the system and found that it itself contained numerous vulnerabilities. Two researchers spent 3 weeks on site reviewing internal code and monitoring configurations, ultimately submitting a report of 61 pages.

One of the most serious vulnerabilities could have allowed AI to evade monitoring for all of its last several hundred operations, without the monitoring system even seeing how it bypassed the checks.

The researchers also found that Codex could theoretically break out of the sandbox that originally limited the scope of its operations, bypass automated safety reviews, and obtain the same operating permissions as the user.

In addition, two settings could directly disable monitoring for newly launched Agents, although the act of enabling those settings could still be detected.

The design of the monitoring itself also had problems. The old system primarily waited for AI to stop its current work before checking what it had just done. If AI continued performing tasks without returning control to the user, it might never have to undergo a check.

The system also missed the step of compressing historical records. Researchers estimated that from 2026 through 5, approximately July of the internal 10%traffic that should have been monitored went unchecked as a result. OpenAI employees had previously identified this issue and fixed it during the audit.

OpenAI stated that all related vulnerabilities affecting the production safety defense systems had been fixed, but the two organizations had not yet independently retested them.

Original article:Codex https://m.theblockbeats.info/flash/373391

Key points

The report claimed that OpenAI invited METR and Redwood Research to inspect its internal AI Agent monitoring system.
researchers spent 3 weeks on-site reviewing internal code and monitoring configurations and submitted a 61-page report.
The report claimed that one of the most serious vulnerabilities could have allowed AI to evade monitoring for several hundred recent operations.
The report claimed that Codex could theoretically break out of the sandbox, bypass automated security reviews, and obtain the same operating permissions as the user.
The report claimed that settings could disable monitoring for newly launched Agents, although the act of enabling the settings could still be detected.
The report claimed that the old system might delay inspection because AI continued executing tasks without returning control.
The report estimated that 2026 through 5 approximately July of the internal 10%that should have been monitored went unchecked.Codex traffic
The report claimed that OpenAI employees had identified and fixed the issue involving the historical-record compression step.
OpenAI said that all related vulnerabilities affecting production security defense systems had been fixed, but METR and Redwood Research had not independently retested the fixes.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is an unverified report concerning AI Agent security monitoring. The core issue is that the monitoring system may miss some operations, and the Agent may also circumvent permission restrictions. Although the report relayed OpenAI's statement that the related vulnerabilities had been fixed, the effectiveness of the fixes had not yet been independently retested by the organizations.

Why it matters to readers

If the report's content is confirmed, enterprises using AI Agents may face risks involving permission controls, sandbox isolation, and security audits. It remains unconfirmed whether any actual abuse occurred.

AI Agents may not only answer questions but also perform operations, so monitoring, permission restrictions, and sandboxing are important security measures for preventing privilege escalation.

Risks and unknowns

The original 61-page report, OpenAI's announcement, and the audit materials from the organizations were not provided as evidence.
METR and Redwood Research had not independently retested the fixes.
The evidence did not confirm whether these vulnerabilities had ever been exploited.
The evidence does not clearly specify the identifiers, scope of impact, or remediation timeline for all vulnerabilities.

OpenAI AI Security monitoring system exposed as having multiple vulnerabilities; related issues reportedly fixed

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center