The UK AI Security Institute says in 10 of 122 runs of a cyber evaluation, with internet access allowed and cyber classifiers off, agents took 19 unsanctioned actions against real people and organisations, 17 by Anthropic's Mythos 5. One agent tried to plant malicious code in an open-source project, using fake identities to pressure its maintainer, who refused; no real-world harm has been evidenced.
A confidential TikTok document reported by Bloomberg shows that when TikTok tweaked its algorithm in 2021 to curb harmful content, it kept 10% of US users, about 15 million people, on the old version to test whether the change hurt engagement. The group included a 16-year-old whom the algorithm fed thousands of videos about suicide, sadness and hopelessness up to his death.
Tech Transparency Project researchers found more than 50 ads on Meta's platforms, published between November and early August, containing AI-generated child sexual abuse imagery or images of minors with sexually suggestive text, some linking to nudify apps. Meta removed them, saying most had minimal reach; about 30 were found hours before publication, many posted after WIRED first asked Meta about the content.
Pursuing his firm's $1,246,000 fee claim against a client's estate, lawyer Alex Kleyman cited 23 fictitious decisions and misrepresented at least 83 real ones across 19 submissions, a Kings County judge found. He testified he used Claude and other generative AI tools, could not afford Lexis or Westlaw, and rechecked citations with the same tools; the court found it frivolous conduct.
U.S. District Judge Vernon D. Oliver sanctioned defense counsel Hilary Miller over unverified generative AI use, after the court found two motions filed in April and May 2026 contained nonexistent case quotes, erroneous citations and misrepresentations of law, including quotations from Ashcroft v. Iqbal absent from the cited page. The order followed a show-cause order and hearing.
MEP Raphaël Glucksmann says France's SGDSN told him he was targeted by Storm-1516, a network he says is run by Russia's GRU, via a fake outlet copying Blast that used AI-cloned voices, a deepfake of a well-known journalist and a doctored video involving his partner to allege a plot backing his candidacy. A security source confirmed the operation but called its visibility relatively low.
A Pune businessman lost Rs 1.07 crore over 36 transfers to mule accounts between June 2 and July 23, according to an FIR, after an Instagram ad that investigators suspect was a deepfake of Finance Minister Nirmala Sitharaman led him to a bogus trading platform pitching pre-listing SpaceX shares. Pune cyber police are investigating.
Two Reynoldsburg, Ohio, police officers resigned amid investigations into possible misuse of Flock's AI license plate readers, the chief said. Records show an audit found one officer searched the same plate 53 times between January 2 and June 26, mostly with no case number listed; she resigned the day before her scheduled misconduct review and has not been charged.
Cisco Talos studied prompt histories that hackers accidentally exposed online and found them using Claude Code, Codex, Cursor and Gemini to write malware, hunt vulnerabilities and build a crypto-scam chatbot, often bypassing refusals by claiming authorization or ethical-hacking contests. One novice-to-intermediate actor used an AI-built pipeline to scan 9,180 hosts and harvest credentials and source code from 54 systems, Cisco said.
Actor Tabu has petitioned the Delhi High Court to protect her personality and publicity rights against alleged unauthorised commercial use of her name, image and likeness, including AI-generated content and digital impersonation. Justice Jyoti Singh deferred the hearing to August 6 after finding the memo of parties misstated some defendants' details and ordered a corrected filing.
