Meta pays $300,000 to a researcher who found a flaw in its AI agent

Meta paid a researcher $300,000 for finding a security flaw in Muse, its AI agent. The flaw exposed conversations and access tokens (digital keys that prove who you are) on test computers Meta staff used before launch. It was the largest award in the Muse bug bounty program.

What happened

Meta says it found the problem about four months before Muse launched on Sept. 8. It says it shut down the affected service and fixed the issue during early internal testing. According to Meta's Nov. 2 disclosure, a check meant to confirm who was asking for data trusted a label added by a local proxy, a middleman that handles web traffic. A system migration added a second proxy, and outside requests could then carry the same label.

Meta says it fixed the flaw and stopped assuming that requests coming from inside the system are safe.

The bigger picture

The bounty program has paid about $1.2 million across 64 valid findings, from more than 630 reports. Meta's disclosure also covers prompt-injection attacks, which trick an AI into following hidden instructions, plus weak credential filtering and flaws in its Android and Mac apps.

What is not known

• How many test machines were exposed.
• What permissions the exposed tokens carried.
• Whether those credentials were revoked.

For any business trying AI agents, the lesson is practical. Ask where an agent stores passwords and what data it can reach, not just what it says.