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:
- AI Agents are here. So are the threats.
May 1, 2025
Agentic applications are programs that leverage AI agents — software designed to autonomously collect data and take actions toward specific objectives — to drive their functionality. As AI agents are becoming more widely adopted in real-world applications, understanding their security implications is critical. This article investigates ways attackers can target agentic applications, presenting nine concrete attack ...
- Hi, robot: Half of all internet traffic now automated
April 16, 2025
Traffic from “bad bots”—those created with malicious intent—first surpassed good bot traffic in 2016, Imperva’s research said, and it’s been getting worse. Bad bots comprised 37% of internet traffic in 2024, up from 32% the year prior. Good bots accounted for just 14% of the internet’s traffic. Bad bots do all kinds of unpleasant things. An ...
- How ToddyCat tried to hide behind AV software
April 7, 2025
To hide their activity in infected systems, APT groups resort to various techniques to bypass defenses. Most of these techniques are well known and detectable by both EPP solutions and EDR threat-monitoring and response tools. In early 2024, while investigating ToddyCat-related incidents, Kaspersky researchers detected a suspicious file named version.dll in the temp directory on multiple ...
- TookPS: DeepSeek isn’t the only game in town
April 2, 2025
In early March, we published a study detailing several malicious campaigns that exploited the popular DeepSeek LLM as a lure. Subsequent telemetry analysis indicated that the TookPS downloader, a malware strain detailed in the article, was not limited to mimicking neural networks. Kaspersky researchers identified fraudulent websites mimic official sources for remote desktop and 3D modeling ...
- Brussels unveils highly-anticipated strategy on defence but fails to outline new funding options
March 19, 2025
The White Paper, released on Wednesday, broadly rehashes the Commission’s ReArm Europe proposal unveiled two weeks ago that could see member states mobilise up to €800 billion for defence over the next four years by notably, pooling orders and procuring equipment together. Capabilities the Commission has identified as priority areas include air and missile defence, artillery ...
- South Korean defense giants bet big on AI-powered military tech
March 12, 2025
South Korean defense companies are ramping up investments in artificial intelligence (AI) software technology to boost unmanned warfare capabilities. The growing role of AI-powered autonomous weapon systems in warfare, particularly highlighted by the Russia-Ukraine war, has made them decisive in determining victory or defeat. Defense firms are making strategic equity investments in AI-focused technology companies to ...