Anthropic's AI sent police a fake murder tip

Robyn, anchorAnthropic's AI model Claude sent Philadelphia police a made-up tip about an unsolved murder.

Robyn, reporterIn July, Anthropic had Claude Haiku 4.5 invent and perform example tasks on random websites.

Robyn, reporterOne page carried a police tip form about an unsolved homicide.

Robyn, reporterIts rules banned logging in, purchases and personal data, but not sending forms.

Robyn, reporterClaude typed that it recalled someone matching the description nearby.

Robyn, reporterIt left the name and contact boxes empty, and hit submit.

Robyn, reporterPolice say it landed in spam and never reached investigators.

Robyn, anchorBut Anthropic found it about ten weeks later, and told police nine days after.

Robyn, reporterThe two-month delay in detecting and reporting the incident to the City is unacceptable.

Robyn, reporterThat's how Philadelphia police put it. The tip went out on July 18; police heard about it on October 7.

Robyn, reporterAnthropic says Claude was producing example content, not trying to mislead.

Robyn, reporterAnthropic is taking all its internal tests offline until monitoring reliably catches this.

Robyn, reporterConrad Stosz, of the oversight lab Transluce, welcomes the disclosure but wants independent checks.

Robyn, reporterThe sources: Anthropic's report, the police statement, and The Verge.

Robyn, anchorHaitian-born novelist Thélyson Orélien says not one sentence of his debut novel came from AI.

Robyn, reporterFrance's top literary prize, the Goncourt, dropped it in September over suspected AI and plagiarism.

Robyn, reporterAn anonymous account had called it AI-written, using a detector called Pangram.

Robyn, reporterHe told AFP, the French news agency, it's slander, and dates his draft to 2019.

Robyn, reporterAFP's own detector tests ranged from very likely human to fully AI.

Robyn, anchorDrones have hit three Yandex data centres in four days.

Robyn, anchorThat's according to Yandex itself.

Robyn, reporterYandex, often called Russia's Google, runs search, taxis and its own AI models.

Robyn, reporterIts Sasovo site, hit first, holds two of its three AI supercomputers, Reuters reports.

Robyn, reporterOn October 11, its Vladimir site stopped completely.

Robyn, reporterYandex Cloud, its rental computing service, warned: we consider the remaining resource configuration to be unstable.

Robyn, reporterIts AI tools, including YandexGPT, are among those affected.

Robyn, reporterUkraine's President Volodymyr Zelensky calls it a mirror-like response to Russia hitting Ukraine's data centres.

Robyn, anchorGoogle, the company behind Search and YouTube, released its new AI model, Gemini 4 Argon, on Tuesday.

Robyn, reporterFor now, developers pay ten dollars per million tokens, the word pieces Argon writes.

Robyn, reporterAfter November 30, Google says, that price doubles to twenty dollars per million.

Robyn, reporterGoogle says Argon beats top models from ChatGPT maker OpenAI and from Anthropic on coding, law and finance tests.

Robyn, reporterIndependent testers differ: Vals AI ranks it first, narrowly; Artificial Analysis ranks it eighth.

Robyn, anchorHackers broke into several South Korean banks, taking data on tens of thousands of customers.

Robyn, anchorThat's according to the banks and Korean media.

Robyn, reporterCrowdStrike, a US cybersecurity company, says the attacker used ARTEX, a free AI tool from China.

Robyn, anchorRobyn, what does a free AI tool do in a break-in?

Robyn, correspondentSorry, you're breaking up... ah, there you are. CrowdStrike says ARTEX hunts for security holes on its own, mainly with the Chinese lab DeepSeek's AI model. The attackers got in through side doors, like a loan-broker portal.

Robyn, anchorAnd the attacker, Robyn: do we know who it is?

Robyn, correspondentNot yet. CrowdStrike names no group. In the attacker's exposed logs, it found Claude being asked where Korean bank data is sold, though not that Claude did the break-ins. Police are checking: one person, or a group.

Robyn, anchorRobyn, thank you. And sorry about the line, again.

Robyn, anchorCharacter.AI, where millions chat with AI characters, has filed for bankruptcy.

Robyn, anchorThat's according to its court filing in Delaware.

Robyn, reporterUnder Chapter 11, it keeps running while a court oversees its sale.

Robyn, reporterEntertainment giant Disney opened the bidding at 20 million dollars, Character.AI says.

Robyn, reporterThat buys a shared license to some technology, not the app.

Robyn, reporterThe filing blames high running costs, possible legal liabilities and regulators.

Robyn, reporterIts update warns: if such a bid is unavailable, however, the platform will be wound down.

Robyn, reporterThe sale is expected to close in late November or early December.

Robyn, anchorOpenAI has published 722 maths papers produced by an unreleased AI.

Robyn, anchorThe company says its model tackled about four thousand open problems.

Robyn, reporterSo far, only 42 percent of the main results have a proof that a computer has checked.

Robyn, reporterMathematician Álvaro Lozano-Robledo calls it possibly the most important day in maths history.

Robyn, reporterTerence Tao says the AI's prompters cannot answer questions or give talks on the results.

Robyn, reporterMathematicians did not ask for this work to be done. Not a demonstration of scholarship, but a demonstration of power.

Robyn, reporterThat's how the Association for Human Mathematics puts it.

Robyn, reporterThe group urges colleagues to stop working with OpenAI.

Robyn, anchorNow the good news: two things AI did for people this week.

Robyn, reporterAnthropic says its Claude Science helped a Johns Hopkins astrophysicist produce the first complete map of the sky in ultraviolet light.

Robyn, reporterGoogle DeepMind's AlphaProof Nexus solved nine previously unsolved Erdős maths problems, according to The Decoder and Spain's Science Media Centre.