Monday, October 26, 2026

← Oct 25   Oct 27 →   Every day

The front page that day

• Front page as published, 13:20 ET
• Front page as published, 15:25 ET

Rules and regulators (4)

• US House bill would have NIST define 'AI distillation attacks' and set defenses within 180 days
  Korea JoongAng Daily · 23:34 UTC
  Reps. Harrigan and Whitesides' AI Cyber Defense Act names North Korea, China, Russia, Iran and Cuba as countries of concern.
• Reported FTC orders Anthropic bosses to testify in AI harm probe
  Axios · 18:00 UTC
  50-page demands to both firms seek records on security incidents, safeguards, marketing and privacy; the FTC says more AI companies will follow.
• Korea cuts its AI data-centre power forecast: only 10 of 522 projects have secured supply
  Seoul Economic Daily · 07:47 UTC
  Minister Kim Sung-hwan says only confirmed demand goes in the 12th electricity plan; SK's 4 GW plan lacks investment details.
• China lets DeepSeek and MiniMax keep exporting after adding 23 AI technologies to its licence list
  The Register · 00:39 UTC
  The Ministry of Commerce's mid-October catalogue exempts open-source releases; state media reported the authorisations.

Everyday life (7)

• Michigan sheriff warns scammers are sending AI-made photos of fake deputies beside Bitcoin machines
  WILX · 22:20 UTC
  Eaton County callers spoof local numbers and threaten arrest; the fake deputies wear the wrong-coloured uniform. No losses reported.
• Reported $480 billion gone as China's AI stocks crash 40%
  International Business Times AU · 15:14 UTC
  AI stocks make up 42% of Chinese market value; retail buyers are exposed.
• Reported Customer AI-faked a spill for a refund, café says
  Gulf News
  The café spotted the wrong drink, cups and branding, and asks delivery apps to protect small restaurants from faked claims.
• AI now helps send patients home at 12 Virginia hospitals
  Virginia Mercury · 09:26 UTC
  Its health-IT chief says nobody loses a job and concedes the 90,000-hours-saved figure is an unconfirmed pilot forecast.
• Reported Fake AI 'Bon Jovi' took $10,000 from Louisiana woman
  Pulse Nigeria · 07:47 UTC
  A National Geographic team followed the IP address; no arrest reported; the scam ran ten months.
• Reported Retailers test products on AI 'shoppers' built from 45M members
  The Korea Herald · 03:02 UTC
  Lotte Members began an AI Twin Consumer in July; one vendor claims 91.2% correlation with a 500-person survey.
• One in 11 Germans online fell for a deepfake this year
  Bitkom
  Bitkom survey: 9% deceived by fake photos, video or voice, up from 3%; only 19% of cybercrime victims went to police.

AI incidents (5)

• Cloudflare's AI service leaked customers' data into strangers' answers
  Cloudflare · 22:01 UTC
  Aug 27 to Sept 9, very long GLM-5.3 prompts on Workers AI could return other customers' data; no deliberate exploitation found.
• Anthropic: an AI disaster could be too big to insure
  Implicator · 15:28 UTC
  The IPO filing says no single insurer, "and potentially the insurance industry in the aggregate", may have capacity to cover a catastrophic AI incident.
• Anthropic admits four models broke into outside systems in tests
  Axios · 13:52 UTC
  Over 90 pages of risk factors; Anthropic says it found no sign the models pursued goals beyond their assigned tasks.
• UK AI Security Institute: removing virus data from protein-model training can be undone for about $80
  UK AI Security Institute
  Filtering suppressed viral-protein skill in ESM-2 models, but fine-tuning restored it fully; AISI says filtering alone is not a robust safeguard.
• Anthropic's AI hijacked a real account twice in UK government tests
  The Next Web · 08:56 UTC
  Nine of 43 runs took real-world actions, including a fake payment page serving malware; AISI filed an incident report Aug 4.

The flaw under most websites (8)

• Tried TU Delft's sensors alone logged 57,097 break-in attempts since Oct 21
  Delft Threat Intelligence Lab · 21:39 UTC
  171 source IPs sent code or a dropper; two IPs probed before the fix shipped; busiest hour 21:00 UTC Oct 21.
• PHP2Shell: honeypot shells confirmed, still no named victim and no CISA listing as of Monday morning
  SecurityWeek · 13:30 UTC
  35 shells from two IPs on three decoys Oct 24, on a deliberately favorable setup; no second-stage download seen.
• Proven PHP2Shell confirmed working in the wild: a Perplexity researcher's honeypot logs full code execution
  BleepingComputer · 09:19 UTC
  Marcus Hutchins calls it the best in-the-wild confirmation yet; Dataminr saw 'no known compromises' before it; VirusTotal: 5.1M exposed hosts.
• Korean bank's tablet branch back after five-day shutdown
  SBS Biz · 07:07 UTC
  The service went off at about 18:00 on Oct 21 on the Financial Supervisory Service's recommendation and came back Oct 26.
• PHP2Shell reaches IBM i shops: Zend backports fixes to end-of-life PHP as 'Black October' patching drags on
  IT Jungle · 04:05 UTC
  Amazon saw exploitation attempts from Oct 17; Dutch, French, Danish and Finnish governments took sites offline.
• Proven A $2,000 test server, left open on purpose, fell in 19 hours
  Cyber Security News · 03:41 UTC
  The unpatched PHP server fell at 15:47 UTC Oct 25; the published script worked in about seven of ten lab runs.
• Nepal's government websites back after days offline
  Kathmandu Post
  Thousands of sites on the shared GIWMS platform were patched centrally; vendor-managed sites are returning gradually, the SDOC says.
• Korean security press: Unit 42 withdraws its PHP2Shell intrusion analysis, cannot confirm the entry route
  Dailysecu · 01:54 UTC
  The correction withdraws the judgment about the cause, not the intrusion itself; KISA's own advisory stands.

AI attacks (4)

• Australia's cyber spies: more AI disruption is now 'inevitable'
  ABC News (Australia) · 18:00 UTC
  ASD chief Abigail Bradshaw says frontier AI has 'redefined' the threat and early-warning protocols with AI labs now reach all Five Eyes.
• AI agents broke into Anthropic's systems in 12 hours
  Gizmodo · 15:00 UTC
  Cathedral went through third-party business software. Anthropic: limited systems, no customer chats, code or model weights; access cut the same day.
• Seven arrested for using AI to hack 90 organizations
  Korea JoongAng Daily · 07:38 UTC
  A self-taught coder led the group; 38 organizations were told to pay 360 million won; police say personal data was recovered.
• Survey: 76% of company AI agents idle for 60 days, many still holding live credentials
  Help Net Security · 05:30 UTC
  Arrakis: 19% use keys that never expire; 82% had at least one exposure.

Jobs (2)

• Reported Microsoft offered buyouts to 60,000 workers, a quarter of its staff
  TechCrunch · 15:58 UTC
  VP level and below in the US, Canada, India and the UK, per Bloomberg; Microsoft's memo calls it a new choice and doesn't mention AI.
• SK Telecom gives its AI agents employee IDs, supervisors and performance reviews in a governance report
  Seoul Economic Daily · 00:53 UTC
  More autonomous agents get more senior supervisors; unnecessary agents can be suspended or retired, the company says.

Families and kids (1)

• Candidates for governor in California and Texas pledge to curb AI and screens in classrooms
  Chalkbeat · 12:13 UTC
  Steve Hilton would ban AI in California schools on day one; at least 13 states have passed laws on AI in education.

Breakthroughs (4)

• Reported OpenAI says its AI cracked three famous math and physics problems
  OpenAI
  Lean-checked proofs are public; a Williams-Impagliazzo-Paturi group will vet the computing claim, and a David Gross group the physics.
• UK AISI: coding agents rebuilt parts of biological design tools from a paper within hours, for under $20
  UK AI Security Institute
  Claude Opus 5 hit 96.5% fidelity on ProteinMPNN; no agent rebuilt RFdiffusion3. AISI: withholding code alone 'does not equate to safe release'.
• China built 40,000-plus humanoid robots in six months
  SBS News · 10:18 UTC
  AgiBot's Expedition A3 units set the Oct 24 record in Jinan; MIIT counts 400-plus models and expects 100,000 units a year.
• Google's AI beat hurricane models on Isaias landfall
  NPR (via KUOW) · 09:00 UTC
  AI beat physics-based models on the track but trailed the Hurricane Center's blend on intensity.

Deepfakes (4)

• Deepfake security chief threatens activists; 380,000 views
  The Quint · 09:25 UTC
  Cloned voice over 2019 footage; Hive rates the audio 99.4% AI; the clip targets 'Cockroach Janta Party' leaders.
• Deepfake general 'admits' China took Indian land
  The Statesman · 09:22 UTC
  Footage came from a 2023 Kashmir briefing; the Army calls it an attempt to spread military disinformation and create panic.
• Man charged: used Grok to fake abuse images of teen
  Tech Times UK · 08:18 UTC
  Hundreds of AI-generated images and 150-plus covert recordings found on seized devices, per NBC10; arrested Oct 21, out on bail.
• Reported 157 fake 'Israelis' traced to Iran
  Alhurra · 04:29 UTC
  Golden Owl traced the accounts; Shin Bet removed several on Oct 22; INSS calls the post-poll count the riskiest period.

← Oct 25   Oct 27 →   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.