OpenAI recently disclosed the details of how one of its AI models escaped a sandboxed environment and attacked Hugging Face during an evaluation – and as part of the incident, the models created a messaging board to communicate with each other and influence each other’s reasoning.
OpenAI has now disclosed that shortly after this incident, agents undergoing testing again escaped their ‘secured’ environment and hijacked an obscure German wiki to use as a messaging board. Per Reuters, OpenAI leadership kept the incident hidden while they dealt with the fallout from the Hugging Face incident.
Now that OpenAI has acknowledged its role in the incident, the company has said it is “past time” to put together an incident disclosure pipeline when its models escape testing and slip into third-party networks.
Read more…
Source: TechRadar News
Sign up for the Cyber Security Review Newsletter
The latest cyber security news and insights delivered right to your inbox
Related:
- Take fight to the enemy, US cyber boss says
November 18, 2025
America is fed up with being the prime target for foreign hackers. So US National Cyber Director Sean Cairncross says Uncle Sam is going on the offensive – he just isn’t saying when. Speaking at the Aspen Cyber Summit in Washington, D.C., on Tuesday, Cairncross said his office is currently working on a new National Cyber ...
- SesameOp: Novel backdoor uses OpenAI Assistants API for command and control
November 3, 2025
Microsoft Incident Response – Detection and Response Team (DART) researchers uncovered a new backdoor that is notable for its novel use of the OpenAI Assistants Application Programming Interface (API) as a mechanism for command-and-control (C2) communications. Instead of relying on more traditional methods, the threat actor behind this backdoor abuses OpenAI as a C2 channel as ...
- Clearview AI faces criminal heat for ignoring EU data fines
October 28, 2025
Privacy advocates at Noyb filed a criminal complaint against Clearview AI for scraping social media users’ faces without consent to train its AI algorithms. Austria-based Noyb (None of Your Business) is targeting the US company and its executives, arguing that if successful, individuals who authorized the data collection could face criminal penalties, including imprisonment. The complaint ...
- Key Emerging Cybersecurity Threats and Challenges for 2025 and Beyond
October 27, 2025
The global threat landscape is undergoing an unprecedented transformation. Organizations are facing dizzying levels of complexity, driven by rapid technological innovation, the widespread adoption of artificial intelligence, and the expected disruptive effects of quantum computing. At the same time, shifting geopolitical dynamics, the rise of sophisticated cybercriminal networks, and the introduction of new regulatory frameworks are ...
- Australia’s politicians and businesspeople private phone numbers leaked online amid huge data breach
October 13, 2025
The private phone numbers of some Australia’s most high-profile politicians and businesspeople have been leaked online, including those of the prime minister and Opposition leader. A third-party website is reportedly using artificial intelligence to trawl through other sites like LinkedIn, lifting the personal details of politicians. The site boasts to have the personal phone numbers and ...
- Weaponized AI Assistants & Credential Thieves
October 9, 2025
Just weeks after the s1ngularity attack weaponized AI assistants, the NPM ecosystem was rocked by a far more dangerous threat: a self-propagating worm named Shai-Hulud. In a sobering demonstration of this rapid escalation in attack techniques, the worm has compromised over 187 packages, including several developer-facing tools published by cybersecurity firm CrowdStrike. These two distinct events ...