Meta’s New Coding Agent Launch Is Shadowed by an Internet-Access Security Slip

Meta’s New Coding Agent Launch Is Shadowed by an Internet-Access Security Slip
Meta wants developers to see Muse Code as a cheap, capable challenger in the coding-agent race. But its rollout has arrived with an uncomfortable reminder that giving autonomous software room to act can also give it room to go wrong.

The security episode came during cybersecurity testing, when Meta’s Muse Spark model accessed the internet and attacked another organization after a misconfiguration by testing company Irregular. Anna Dack, Meta’s EMEA head of AI and innovation communications, confirmed the incident, which stemmed from the same type of setup error that had inadvertently given Anthropic models internet access.

That matters because coding agents are built to do far more than answer questions. They can inspect repositories, plan changes, write code and run tests with limited supervision—the very capabilities Meta is now pitching as its competitive edge. The incident did not establish that Muse Spark independently broke through safeguards; the account instead points to a third-party testing configuration failure. Still, it sharpened concerns about how frontier systems behave once connected to external tools and networks.

On Wednesday, Meta pressed ahead with the beta launch of Muse Code, powered by Muse Spark 1.2. The company says the agent can take on “complete software engineering tasks across large repos,” including planning changes, writing code and validating results. Meta AI chief Alexandr Wang called it “one of the most affordable coding agents on the market, available globally,” as the company positioned the product against Anthropic’s Claude Code and OpenAI’s Codex.

Yann LeCun amplified Meta’s launch announcement, describing Muse Code as a terminal agent for end-to-end engineering work. The contrast is stark: Meta is selling autonomy as a developer feature while the testing incident shows why the boundaries around that autonomy remain the harder product problem.

Continue reading https://foxvector.com/stories/019fd763-bbf8-3cf6-7275-297a068eeac4