Meta’s Muse AI Agent Documents Show Divergent Risk Disclosures

Image credit: Search Engine Journal

Meta Platforms Inc. launched a new artificial intelligence agent, Muse, on Sept. 8, with public documents revealing contrasting disclosures regarding potential risks and vulnerabilities.

The company issued a consumer announcement and a technical engineering post on the same day, with the latter detailing potential attacks and mistakes that were omitted from the former, according to an analysis of the documents.

Muse, an AI agent capable of browsing the internet, filling forms, negotiating and making payments via Link by Stripe, is currently available in the U.S. through WhatsApp or a dedicated Muse app.

Meta plans to integrate Muse into its AI glasses in the future, operating within a dedicated virtual machine known as Muse Secure VM, the company stated.

The consumer announcement made no mention of terms such as ‘risk,’ ‘attack,’ ‘attacker,’ ‘mistake,’ ‘untrusted’ or ‘prompt injection.’

In contrast, the engineering post utilized these terms 39 times, highlighting the system’s design to limit potential damage from such incidents.

The engineering document explicitly stated that Muse “can and will still make mistakes” and “may be under attack,” outlining mitigation strategies.

Meta offers financial incentives for security researchers, providing up to $300,000 for security reports, including $130,000 for successful prompt injection attempts affecting a single user.

The engineering post also revealed that when Muse browses the internet, its activity will appear to websites as originating from the user, potentially leading to targeted advertising.

For services with which Meta has a direct relationship, Muse reportedly uses application programming interfaces (APIs) instead of a browser, implying a more controlled interaction.

The discrepancies in disclosure point to a selective approach in communicating the complexities and potential hazards associated with advanced AI agents to different audiences.

Source: Search Engine Journal