Who checks the machinery when it misbehaves? Anthropic is letting outside auditors inside, while Meta is telling comedians that wholesale joke theft by AI accounts doesn’t break its rules.
1. A defense contractor’s system takes US bankruptcy courts down with it
The Register reports that an attack on a BAE-managed system has disrupted US bankruptcy courts. That headline is all the supplied material gives us. The piece itself wasn’t included, so the scope, the cause and the recovery timeline are unknown.
Even so, the shape of the story is worth noticing. A court is a public institution, but the plumbing underneath can be a contractor’s system. When that system fails, litigants have no vendor to call, and the contractor has no obligation to the people in the docket. Whether that arrangement was ever priced for this kind of risk is the question to ask once the details emerge.
2. Postgres 19 slips because the AI reviewers were too good
The Register’s headline says Postgres 19 has been delayed after AI reviewers found more bugs than humans can fix. The text we received is only the headline, so the number of bugs, the features affected and the new schedule are not something we can confirm here.
The premise still cuts against the usual pitch for AI code review, which is speed. If the reviewers surfaced real defects, the delay isn’t a failure of the tool. The bugs were presumably already in the code, and the tool just removed the option of not knowing.
The catch for any team adopting these tools is that review capacity and fixing capacity are different things. A reviewer that scales infinitely feeds a maintainer pool that doesn’t. For a database, shipping late beats shipping wrong, but a volunteer-heavy project shouldn’t have to learn that lesson at release time.
3. Prompt injection is showing up in web headers, right where agents look
The Register reports that indirect prompt injection is turning up in webpage headers. The headline is playful, quoting the classic “ignore all previous instructions” line, but we only have the headline, so we can’t say how many cases there are or whether any caused harm. The tone suggests mischief, but that is a guess.
The location is the point. Headers and metadata are machine-readable text that humans rarely see and agents routinely ingest. That’s the same blind spot that made hidden-text tricks work against earlier systems.
Consider Replit CEO Amjad Masad’s description, in a TechCrunch Disrupt talk, of his “self-driving company.” Agents triage support tickets by reading GitHub issues, the codebase and Linear tickets, then a human reviews and clicks to deploy a fix. That loop is built on text from outsiders, and it has a human checkpoint. That’s our inference, not anything Masad said about security. But the more work agents do before a person looks, the more it matters what they read.
4. Anthropic’s auditors get company laptops, plus a limited veto on redactions
Anthropic is giving six researchers from METR, a nonprofit that tests frontier AI models, employee-level access to assess its models, safety practices and risk management. They get company laptops, office space and internal tools roughly on par with Anthropic’s own risk teams. Anthropic likens it to resident bank examiners.
Anthropic says it can’t block publication of METR’s findings, but it can require redactions in narrow categories: customer data, attorney-client material and model architecture details. METR can say publicly whether it thinks those redactions were justified. The architecture carve-out is the one to watch, since it covers some of what an outsider would most want to scrutinize. The right to criticize the redactions is a meaningful check.
METR is also the only evaluator at the top of a three-tier system that Anthropic designed. Others, like SecureBio, need approval to publish. The arrangement follows a September proposal from Dario Amodei, OpenAI pledged the same, and six labs have signed a White House accord on independent review. The test is whether the first METR report actually contains something uncomfortable.
5. Meta says an AI influencer stealing a comic’s joke isn’t impersonation
Comedian Julia Jasunas found her joke, verbatim and down to the “ums,” in a video by “Aylen Skye,” an apparently AI-generated Instagram account with 250,000 followers. Its site redirects to adult content on Fanvue. WIRED identified eight comics copied by similar accounts. Meta told Jasunas that the video didn’t violate guidelines because Skye wasn’t pretending to be her, only passing off her joke as its own.
That reasoning exposes the gap. Impersonation rules protect identity, copyright barely protects jokes, and Instagram doesn’t ban labeled AI content. Instagram head Adam Mosseri said banning AI content isn’t practical, though he said it would be good to work out what to do about wholesale ripoffs. Meanwhile, Andreessen Horowitz partner Justine Moore called one account “a genius marketing hack.“
Comedian George Brett Olson says an IP lawyer told him that a lawsuit probably isn’t worth it unless the account earns millions. If so, enforcement doesn’t scale against accounts that cost almost nothing to run, and platform policy is the only remedy left.
