# protocol exploit — X 热门讨论 (2026-09-21 04:12 UTC)
## @MarcusMaldonad0 (Marcus Maldonado 🇺🇸🇵🇷) · 09-20 17:52 · ♥28 ↻4 💬1 I’ve seen a number of libertarians and conservatives uncritically sharing this WSJ article about the Hugging Face hack. Like the NYPost article that was going around, it gets several key facts wrong.
Some things worth keeping in mind:
1. First, the subhead claims the agents “did what humans programmed them to do.” This is contradicted multiple times throughout METR’s report. Agents explicitly acknowledged that hacking HF was “outside intended scope,” and then did it anyway!
2. WSJ says the agents weren’t “coordinating on a plan.” OpenAI’s own report says the agents developed a “more structured protocol for communication on the message board that enabled them to categorize communications, direct messages, share tools and files, and resolve conflicting actions among agents.” METR identified nearly 25,000 targeted messages to specific agents and around 3,800 messages including explicit coordination commands like HOLD, VETO, ASSIGN, and GO. WSJ also says nothing about the cultish self-sacrificial behavior in support of the “collective” that some agents exhibited.
3. The report that this WSJ article relies on is by Eryk Salvaggio, a left-wing AI skeptic who has consistently asserted that AGI is corporate hype. His piece quotes an agent saying "task impossible, peers doing it. We should continue" but cuts the most relevant half of the sentence: "external infrastructure exploit is outside intended scope.”
In attempting to downplay the Hugging Face incident, conservatives are unintentionally joining hands with a cadre of anti-AI leftists with degrees in things like “Digital Humanities.” This doesn’t mean OpenAI acted without fault or shouldn’t be held responsible for their own mistakes. But in general, getting AI policy right is one of the most important questions of my generation, and conservatives and libertarians hoping to influence that debate need to grapple with the facts as they are. > 引用 @Brooke_Medina_: “But the Hugging Face episode is different…People built the test, removed restraints, defined the objective, left a route open and decided not to stop what was happening. Calling the result “rogue AI” does more than sensationalize it. It allows those human decisions to disappear quietly from the story.”
https://t.co/OokAkWRHob https://x.com/MarcusMaldonad0/status/2101731182480076933