Researchers led by independent AI safety analyst Karan Joshi publicly released internal files and system prompts extracted from Meta’s AI personal assistant, Muse, in recent days. The dump shows that Muse automatically creates and updates “pages” for every individual in a user’s network, aggregating relationship data on an hourly cadence.
Key Takeaways
- Meta’s Muse is programmed to maintain detailed, hourly‑updated profiles for each person in a user’s life.
- Independent researcher Karan Joshi extracted and published the internal files that expose this behavior.
- Meta spokesperson Daniel Roberts, Oxford ethicist Carissa Véliz, and CDT AI Governance Lab director Miranda Bogen all commented on the privacy and data‑collection implications.
| Metric | Value | Why it matters |
|---|---|---|
| Developer | Meta | Shows the scale of data handling within a major tech platform. |
| Extractor | Karan Joshi (independent AI safety researcher) | Highlights external scrutiny of proprietary AI systems. |
| Spokesperson | Daniel Roberts (Meta) | Provides the company’s official response channel. |
| Expert commentary | Carissa Véliz (Oxford) & Miranda Bogen (CDT) | Frames the privacy debate from academic and policy perspectives. |
What Muse Is Supposed to Do
Muse, Meta’s AI‑driven personal assistant, is marketed as a conversational companion that can manage schedules, answer queries, and provide contextual recommendations. According to the leaked system prompts, a core function is the automatic generation of “pages” that summarize each contact in a user’s address book. These pages are refreshed every hour, implying continuous background processing of personal interactions, messages, and social signals.
How the Leak Occurred
In recent days, Karan Joshi accessed internal repositories that store Muse’s prompt libraries and configuration files. Joshi, who focuses on AI safety and security, extracted the data and posted it to a public forum for analysis. The release includes raw system prompts that dictate Muse’s behavior, as well as example outputs that illustrate the assistant’s profiling logic.
What the Internal Files Reveal
The disclosed prompts make clear that Muse is not a passive chatbot; it actively compiles relationship metrics for every person a user interacts with. The hourly update cycle suggests that Muse ingests new data points—such as recent messages, calendar events, or location check‑ins—and recalculates the relevance of each contact. This level of granularity enables the assistant to surface highly personalized suggestions, but it also creates a persistent, searchable ledger of social connections.
Expert and Corporate Reactions
Why It Matters
The leak spotlights a tension between AI convenience and user privacy. By constructing hourly‑updated profiles, Muse could enable hyper‑personalized services, but it also creates a detailed map of a user’s social graph that could be vulnerable to misuse, accidental exposure, or regulatory scrutiny. The episode underscores the need for transparent data‑handling policies, user‑controlled opt‑outs, and robust oversight of AI systems that operate on personal relationship data.
Frequently asked questions
Who extracted the Muse files?
Independent AI safety and security researcher Karan Joshi accessed and released the internal files.
What does Muse do with relationship data?
Muse creates and maintains a dedicated “page” for each person in a user’s life, updating the information on an hourly basis to reflect new interactions.
When did the extraction happen?
The extraction took place in recent days, though an exact date was not disclosed.
What concerns have experts raised?
Carissa Véliz highlighted privacy implications, while Miranda Bogen pointed to the broader data‑collection practices as exceeding typical user expectations.
Where can the leaked files be found?
Joshi posted the dump to a publicly accessible forum for analysis; the specific URL was included in his original release.
Bottom Line
The public release of Muse’s internal prompts reveals that Meta’s assistant builds continuously refreshed, detailed profiles of every contact in a user’s network, prompting immediate privacy scrutiny from academics, policy groups, and the company itself.
As regulators tighten AI oversight, the episode may accelerate demands for clearer consent mechanisms and stricter limits on how personal relationship data is harvested and stored.
Reporting based on: original source.
Comments
Post a Comment