Meta's Muse Bug Bounty Discoveries: Payouts Exceed $1.2 Million

Meta has distributed approximately $1.2 million in rewards through its bug bounty program for the Muse AI agent, highlighting critical vulnerabilities and security improvements.

Meta’s Bug Bounty Program

Meta’s bug bounty initiative, which operated privately from April and transitioned to public access after Muse’s launch on September 8, has awarded significant payouts for verified vulnerabilities. The program, detailed in a blog post by Meta’s David Singleton, received 630 reports, 64 of which were validated.

Critical Vulnerability Identified

A $300,000 payout was issued for a critical vulnerability identified before the system’s public release. A researcher demonstrated access to internal communications and authentication tokens on pre-release virtual machines used during Meta’s internal testing phase. The flaw stemmed from an authentication mechanism that mistakenly validated headers from a local HTTPS proxy as evidence of trusted origin.

“During a system migration, an additional proxy layer was introduced, allowing external requests to bypass security checks by carrying the same trusted headers,” according to Meta’s blog post.

Meta addressed the issue by temporarily suspending the service, patching the vulnerability, and reconfiguring the system to no longer treat local environments as inherently trusted.

Additional Rewards for Security Flaws

A separate $90,000 reward was awarded for a report highlighting a malicious web page that triggered Muse to mimic status messages from its browsing tool. The attack exploited the agent’s ability to execute actions based on unverified inputs, prompting Meta to implement enhanced detection mechanisms.

Another $90,000 was paid for a vulnerability where specific formatting techniques could bypass Muse’s one-time passcode redaction filters. A $50,000 bounty was issued for an Android app flaw enabling unauthorized access to privileged web views, while a $1,000 reward was given for a Mac app vulnerability allowing code execution via the auto-update feature.

Reward Structure and Impact

Meta’s bug bounty program offers up to $300,000 per verified report, exceeding rewards from other major AI developers. Prompt injection attacks targeting individual users can yield up to $130,000, while widespread compromises may qualify for as much as $250,000. Testing confirmed that updated classifiers now block this specific attack vector and 100 related variants.

Conclusion

Meta’s proactive approach to security through its bug bounty program underscores the importance of addressing vulnerabilities in AI systems. By incentivizing researchers and swiftly resolving issues, Meta continues to strengthen the security of its Muse AI agent.