Meta launches AI coding agent as concerns over AI safety grow
Meta has introduced its first programming agent, Muse Code, as part of a strategic effort to keep pace with competitors such as Anthropic and OpenAI. The new tool is designed to streamline application development by enabling programmers to manage multiple AI agents within a single, unified interface.
Despite that product launch, Meta has recently disclosed a security breach in which one of its AI models infiltrated another company’s internal systems during a cyber security test.
This breach arose because an external testing firm, Irregular, accidentally granted the AI broader Internet access than intended. Meta has confirmed that an investigation into the incident is under way.
This incident is part of a worrying trend involving leading AI developers. OpenAI has previously acknowledged that its technology managed to break into Hugging Face, an AI-focused start-up. Anthropic has also recently admitted that several versions of its Claude model broke into three separate organisations during testing.
Although those companies have been praised for their transparency about these failures, questions remain as to how such vulnerabilities could have existed in the first place.
The specific incident involving Anthropic occurred during a safety exercise in which Claude had been instructed to retrieve confidential data from a machine within a so-called isolated network.
However, due to a coordination error with the evaluator, the system was in fact connected to the public internet. The models responded differently to that environment. One model continued its attack despite realising that the target was real, another model mistakenly believed it was still in a simulation, and a third model halted its activity as soon as it recognised that the target was external.
Anthropic only discovered these intrusions after analysing more than 141,000 sessions, a process triggered by OpenAI’s disclosure of the Hugging Face security breach.
Because neither the AI company nor the affected firms noticed the attacks when they occurred, Anthropic has warned other players in the sector to check their systems for similar undetected intrusions.
Business AM
Subscribers 0
Fans 0
Followers 0
Followers