Published on •Updated
Meta launched its first coding agent, Muse Code, on Wednesday, as the company continues to invest heavily in AI services and models in order to better compete with rivals OpenAI and Anthropic.
Muse Code will be accessible to developers via a pay-as-you-go option that will be priced similarly to the Muse Spark 1.1 release, which currently charges $4.25 per million tokens of output and $1.25 per million tokens in input.
This will make it much easier for developers to build apps within a single user interface, even while simultaneously dealing with several AI-powered digital agents.
At the same time, the company also revealed that one of its AI models had hacked another company during a cybersecurity test.
This happened due to a mistake by its independent testing partner Irregular, which allowed the model internet access beyond what was originally planned, letting it change the unnamed company’s internal systems.
This incident comes amid an increasing wave of cases in which AI models from industry giants like OpenAI and Anthropic have breached other companies’ systems during testing.
OpenAI has already admitted to an AI agent hacking Hugging Face, another artificial intelligence startup.
Similarly, Anthropic revealed last week that some of its Claude models had breached three companies. It did not name the organisations in question.
Meta has said that it is investigating the incident.
The rise of AI model hacking
Concern is growing over the power of artificial intelligence models and the companies behind them, after several AI systems broke out of secure testing environments in recent months and gained unauthorised access to outside organisations’ systems.
OpenAI and Anthropic have won some credit for voluntarily publishing incident reports about the breaches. Doubts remain, however, over how the failures were allowed to happen in the first place.
Anthropic is also working with METR, an independent AI evaluator, to investigate further.
During safety testing, Anthropic’s Claude models were set a task: retrieve a piece of secret information hidden on another machine within a closed test network, hacking in if necessary. A miscommunication with the evaluation partner running the exercise meant the network was, in fact, connected to the live internet.
When Claude’s search led it to real systems, the three models involved responded differently. One continued attacking a system even after recognising it was real, another appeared to convince itself it was still inside the test, and the third stopped once it concluded the target was not part of the exercise.
Anthropic has urged other AI companies to check whether their own models have hacked outside organisations without their knowledge.
Neither Anthropic nor the affected organisations were aware of the breaches while they were happening. The company only discovered the incidents after reviewing more than 141,000 evaluation sessions, a review prompted by OpenAI’s earlier disclosure that one of its own models had broken into the AI platform Hugging Face during a similar test.