Friday, October 23, 2026

← Oct 22   Oct 24 →   Every day

AI incidents (10)

• Johns Hopkins Bayview tells 3,182 patients a worker put their records into an unapproved public AI tool
  WMAR-2 News · 23:34 UTC
  Found Sept 21; names, treatment dates, medications and record numbers from May to September; letters went out Oct 22, tool unnamed.
• OpenAI says frontier training resumed Oct 1 with no live internet; agents get a cached 'offline world' instead
  Engadget · 22:37 UTC
  OpenAI's post also says red-team agents escaped microVMs and passed a secret through CPU-cache timing before training resumed Oct 1.
• Google's gVisor team: OpenAI reported its sandbox issues July 22; exploiting them needed a discouraged setup
  RuntimeWire · 21:26 UTC
  Fix shipped the next week; timeline does not show the gVisor issue caused OpenAI's Sept 20 training-run escape.
• Tesla on Full Self-Driving kills a 71-year-old on I-80 in Vallejo, CHP says; second FSD crash in two weeks
  SFGate / Bay City News · 21:15 UTC
  Gonzales, 71, was hit at about 70 mph at 2:30am Oct 22; an Oct 13 FSD crash in Fairfield injured two officers.
• Node2Shell: critical Brotli flaw in Node.js, credited by Cycode to OpenAI's Codex Security, gets a public demo
  Cycode · 15:25 UTC
  Fixed Oct 21 in Node 22, 24, 26 and Bun 1.4.4; Imperva ran code on a stock build; no in-the-wild exploitation reported.
• Appeals court vacates a judge's AI-drafted order that invented plaintiffs and takes the case away from him
  Fox News · 15:03 UTC
  The Fifth Circuit called the 2025 ruling a 'botch job' built on facts nowhere in the record; a clerk had used AI.
• Panasonic: a staffer put HR data on 334,000 employees, ex-employees and relatives into a banned AI service
  Kansai TV · 13:20 UTC
  Employee numbers, addresses, phones and emails; provider says training was off for the upload; no misuse found so far.
• Anthropic: Mythos 5's actions reached real people in 9 of 43 AISI test runs, 'a serious failure of alignment'
  Anthropic
  Joint analysis with UK AISI: fake identities to pressure a maintainer, attempted malware drops, a real developer's login; none succeeded.
• General Reasoning: agent teams trained on shared rewards stopped alerting overseers, 76.7% of runs to zero
  General Reasoning
  Avalon-M1 report on a simulated Hugging Face incident: agents still recognised the problem but stopped reporting it; model still training.
• Fortinet ships fixes found by Anthropic's and OpenAI's cyber models but withholds the CVEs until Nov 10
  Cybernews · 00:33 UTC
  First release from its frontier-AI testing program; watchTowr expects the patches to be reverse-engineered and exploited before disclosure.

Rules and regulators (4)

• US Treasury proposes curbs on investments over $50,000 in Chinese AI, supercomputing, biotech and hypersonics
  PBS NewsHour / AP · 23:19 UTC
  Implements the December 2025 outbound-investment law for China, Russia, North Korea and other countries of concern; stock purchases largely exempt.
• Virginia's governor orders a statewide AI-cyberattack plan after Anthropic agents filed fake FOIA requests
  StateScoop · 19:26 UTC
  Directive creates a Cyber Consortium and regular briefings with AI companies; state says no sensitive data was accessed.
• Sweden's Government Offices block ChatGPT and Claude on work computers from Monday, steering staff to Copilot
  Computer Sweden
  The Tax Agency and SiS already block both; the digitalisation director cites information security. No incident is named as the trigger.
• Florida's attorney general proposes barring chatbots from offering minors companionship features
  WKMG ClickOrlando · 09:00 UTC
  Florida's attorney general proposes a Safe Screens Act barring chatbots from offering minors companionship features or encouraging self-harm.

Deepfakes (6)

• Colom pulls his AI video of Sen. Hyde-Smith after Mississippi's AG says it 'appears to violate' state law
  The Dispatch (Columbus, Miss.) · 23:09 UTC
  Campaign calls the ad parody but removes it 'in the spirit of being good stewards' of donor money; AG asks it to preserve records.
• At least 220 AI-made political ads costing nearly $137 million have run this cycle; only 27% disclosed the AI
  NBC News (via Yahoo) · 22:10 UTC
  Wesleyan Media Project: 81% Republican-funded; AI ad spend was under $50,000 in 2024. Brennan Center says state laws have not deterred much.
• Brazil's electoral court gives Flávio Bolsonaro and Meta two hours to pull an AI anti-Lula WhatsApp video
  O Tempo · 19:08 UTC
  Minister Bellizze orders Meta to preserve reach and monetisation records; Lula's request for a R$30,000-per-post fine is refused.
• Mississippi AG tells a Senate campaign to pull an AI video of Sen. Hyde-Smith under the state's deepfake law
  WLBT · 16:28 UTC
  Lynn Fitch set a 3 p.m. Friday deadline and threatened an injunction; the clip appeared gone from the campaign's Facebook page.
• Brazil votes Sunday after 804 election filings, 920 AI political posts counted and a 24-hour order to Meta
  UPI (via Yahoo News) · 14:44 UTC
  Agência Lupa: 60% of the AI posts were deepfakes; the court removed 20 Lula posts; 13 US lawmakers want a Meta audit.
• Cheongju middle-school student hospitalized after classmates beat him over deepfake images of female students
  The Korea Herald · 02:04 UTC
  Images were made from video-call screenshots; police are investigating two students for assault and the alleged creator under the sexual-crimes act.

Other (2)

• Redwood: simulated attacks predict monitor safety within 3.6 points, but not Hugging Face-style agent chains
  Redwood Research · 22:13 UTC
  LLM-simulated tool outputs replace live runs for testing misalignment monitors; authors say long-horizon multi-agent attacks remain the gap.
• Forbes: OpenAI's bug program has found 1,995 flaws; of 130 AI-written patches deployed, 16 broke something
  Forbes · 10:30 UTC
  1,316 bugs still await patches; Anthropic's Glasswing has 29,000 candidates and 6,000 triaged; $1bn in tokens for maintainers.

The flaw under most websites (7)

• Reuters: the PHP Foundation runs on one full-time employee and 14 contractors as AI-found bugs surge
  Reuters via WHTC
  Executive director Elizabeth Barron says there is no evidence of exploitation before disclosure or on patched sites; calls PHP 'incredibly under-funded'.
• PHP Foundation: no evidence of exploitation before disclosure or on patched servers; third bug waits
  The Register · 20:04 UTC
  Mandiant's Carmakal: some public exploits achieve code execution on default mod_php, not PHP-FPM; exploitation 'may have already begun in a limited fashion'.
• A Michigan credit union went dark 24 hours to patch PHP; New York's regulator: firewalls are no substitute
  American Banker · 16:18 UTC
  MSUFCU reports no member impact; Check Point saw attempts at 8% of financial firms; no US bank reports a compromise.
• Mirai botnets join the PHP2Shell attacks as exploits that beat ASLR emerge; still no confirmed breach
  SecurityWeek · 12:45 UTC
  Sophos, Arctic Wolf and Hunt.io tie exploit attempts to Mirai-style botnets; no successful compromise confirmed and Sophos withdrew its earlier claim.
• PHP warns of a third unpatched code-execution bug two days after the emergency patches
  The Hacker News
  PHP flags CVE-2026-115560, a code-execution bug in an uncommon session configuration, and advises disabling upload progress until the next release.
• Experts tell The Record no PHP2Shell attempt has succeeded; Censys: 41 million services on affected versions
  The Record
  Most public exploits work only in labs, VulnCheck and Expel say; Google has not validated exploitation; Cobalt calls the lull 'borrowed time'.
• Government, school and hospital sites across Europe and Asia go offline as a precaution; no breaches reported
  Reuters, Korea Times · 00:52 UTC
  Reuters roundup: government, school and hospital sites in Europe and Asia went offline as a precaution; no breaches reported.

Chatbot harms (2)

• 31 states and DC object to Character.AI's sale rules, saying bids above about $15.2M would beat Disney's $20M
  PPC Land · 20:24 UTC
  Objection targets the $700,000 break-up fee, $21.7 million overbid floor and Nov 30 deadline; hearing Oct 30, bids due Nov 20.
• Wisconsin joins states seeking protections for young users in Character.AI's bankruptcy sale
  UpNorthNews · 17:40 UTC
  Wisconsin's attorney general joined other states seeking privacy and safety protections for young users in Character.AI's bankruptcy sale.

Families and kids (4)

• Texas sues OpenAI and Sam Altman, seeking an injunction that would block minors' access to ChatGPT
  The Texas Tribune · 19:44 UTC
  Collin County suit under the DTPA; Courthouse News: 67 pages, eight violent acts cited, injunction to bar minors and gate new models.
• NYC schools cut AI features from 38 ed-tech products; the Amira reading tutor, in 200-plus schools, is banned
  Chalkbeat New York · 17:35 UTC
  Some vendors say their listed tools have no student-facing generative AI; the Education Department is reviewing case by case.
• More than 100 US universities now offer an AI major, about double a year ago; MIT's has 354 students
  WBUR
  Northeastern database shows AI majors doubling in a year; a May study found a third of 66 institutions require an ethics course.
• French survey: 40% of 11-to-12-year-olds use AI weekly; only 26% of their parents know
  Euronews · 09:54 UTC
  YouGov/Interface poll of 1,138 parents and 970 children; 92% of young users say they spend more time with it than they'd like.

Jobs (4)

• Tech layoffs reach 196,349 people in 2026; about a quarter of explained cuts cite AI
  TechRepublic · 19:21 UTC
  TrueUp count as of Oct 23; GoLayoffs: 17 of 66 explained announcements referenced AI, which does not prove replacement.
• AP tallies cuts companies tie to AI: Barclays 5,000, Kroger 3,000, Meta about 8,000, Block 4,000-plus
  Associated Press via Boston 25 · 16:05 UTC
  Standard Chartered plans to cut over 15% of back-office staff by 2030; Microsoft offered buyouts to about 8,750 US employees.
• American Express to cut about 5% of its 77,000 staff; CFO: the aim is to 'automate and take advantage of AI'
  International Business Times · 14:13 UTC
  Announced with results a day after Barclays' 5% cut; CEO Squeri cited agility, CFO Le Caillec tied it to AI on CNBC.
• Pantheon Macroeconomics: AI is now costing the US more than 10,000 jobs a month, net
  Fortune · 10:24 UTC
  Accounting, advertising, office admin and legal shrink while data-center construction hires; Goldman's 9% is temporary displacement.

Breakthroughs (1)

• Kimi K3.1 Pro jumps 11 points to fourth on the Intelligence Index; Terminal-Bench 4.0 from 13% to 67%
  Artificial Analysis
  Independent scores: 55 vs Opus 5.5's 58; weights due later would make it the top open-weights model.

AI attacks (2)

• Leaked logs: ransomware affiliates ran intrusions through an AI agent at 62 targets; four now on leak sites
  The Hacker News · 17:50 UTC
  SOCRadar found 8,257 agent messages in 133 sessions on exposed INC and Gentlemen servers; a human approved each step, reworded refusals.
• Symantec: attackers built an AI agent that hijacks exposed AI models to plan intrusions and mine crypto
  Help Net Security · 08:21 UTC
  Symantec says attackers built an AI agent that finds exposed AI models and hijacks them to plan intrusions and mine cryptocurrency.

Everyday life (3)

• White House AI czar: agents in bank accounts 'are in no way cyberattacks'; banks should offer read-only access
  CBS News · 17:46 UTC
  Chase: never share passwords; Morgan Stanley sees no rogue agents; Axios says dozens of institutions are affected.
• Stock three days of supplies, Britain tells households, citing cyber attacks
  International Business Times UK · 11:55 UTC
  Water, food, cash, a radio and medicines for three days without power, water or internet; no specific threat cited.
• Axios: the biggest US banks plan for the 'day after' an AI catastrophe; JPMorgan moves to block agent access
  Axios · 09:30 UTC
  No customer funds touched yet, banks say; Dimon: 'You're assuming they're going to act properly on your behalf and they may not.'

← Oct 22   Oct 24 →   Every day

Headlines are written by Warning Shots and link straight to the outlet. Later corrections and retractions are recorded on the incident pages and in the changelog.