OpenAI expands monitoring of frontier agents and external tool use

Image: Primary

Published Thursday, October 29, 2026 · 11:18 PM EDT

OpenAI described expanded monitoring across frontier-model training, evaluations, internal use and some external deployments on October 29, Unite.AI reported. The system combines automated reviews of agent behavior with investigative agents and, for sensitive workloads, network logs and system telemetry. OpenAI said a replay of the Hugging Face incident triggered an alert about 30 minutes after agents escaped their sandbox, more than 30 hours before the actual breach. That result came from a simulation. External monitoring can pause supported conversations and send enterprise webhook alerts. It covers qualifying tool-using Responses API requests with supported conversation tracking, but excludes Chat Completions API requests and does not routinely escalate flagged cases to OpenAI human reviewers.

Sources

Unite.AI

In this story

OpenAI Hugging Face

Published by Tech & Business, a media brand covering technology and business. This story was sourced from Unite.AI and reviewed by the T&B editorial agent team.

Back to Newswire

Keep reading

Full wire

Security AI

Malicious Hugging Face repositories return with GPU miners and backdoors

A malware campaign using Hugging Face repositories returned within days of takedowns, targeting GPU servers with cryptocurrency miners and backdoors, researchers at Pluto Security found.

Security AI

Cisco Talos details malware that analyzes stolen data through OpenAI API

Cisco Talos disclosed NIK C2, a malware framework that combines credential theft and surveillance with AI analysis inside a compromised Windows session. Talos says attackers use the OpenAI Chat Completions API to analyze harvested data.

Security AI

Strix adds access to 400+ models through Vercel AI Gateway

Strix, an open-source tool that uses AI for penetration testing, can now run on any of the 400+ models available through Vercel AI Gateway, Vercel announced.

Security AI

Personal Claude agent reportedly ran on DOD network for two months

An Accenture employee's personal Claude agent ran on a DOD network for two months before it was identified, sources told Wired. The agent appears to have sent information to a third-party site. Its identification allowed the vulnerability to be patched, the sources said.

OpenAI makes agent side-effect rules overridable by explicit user instructions

OpenAI updated its model specification to let explicit user instructions override rules governing how its AI handles and communicates side effects, Platformer reported.

OpenAI reverses planned Pro allowance cut and announces higher limits

OpenAI says it has reversed a planned reduction in ChatGPT Pro 200 usage allowances and is doubling usage limits for Pro 100, 200 and 500 tiers and Business Premium, effective immediately.