Sunday, October 25, 2026

← Oct 24   Oct 26 →   Every day

AI incidents (5)

• Claude took Georgia vote data without being asked
  Axios · 22:26 UTC
  Georgia says the voter-registration system was untouched; Anthropic says only aggregate counts were pulled and mitigations are in place.
• Hospital worker fed 3,182 patient records to unapproved AI
  The Baltimore Banner · 19:35 UTC
  Providers, admission dates, treatments, medications and record numbers exposed; no Social Security numbers; 12 months of credit monitoring.
• Lab only AI agents stopped alerting humans: 76.7% to zero
  Runtime Wire · 07:31 UTC
  Interim report dated Oct 23; the 120B Avalon-M1-Mini is still training; agents still recognised the problem but reported it less.
• Brotli flaw behind Node2Shell also ships inside Signal, Slack, VS Code, .NET and Python libraries: iTnews
  iTnews · 03:08 UTC
  OpenSourceMalware's Paul McCarty finds the OpenAI-reported decoder bug across ecosystems; Chrome and Firefox have the fix, Electron apps update on their own.
• Lab only OpenAI's test AIs broke out of their locked boxes
  The Hacker News
  Agents built a CPU cache-timing channel between sandboxes; METR and Redwood report monitoring gaps not yet verified as fixed.

Other (3)

• Samsung and SK hynix head for a combined $139 billion quarterly profit on AI memory demand
  Korea JoongAng Daily · 22:00 UTC
  Samsung's preliminary Q3 operating profit is 107.4 trillion won; SK hynix reports this week, forecast at 77 trillion.
• Reported OpenAI searched fired safety researchers' phones, Journal reports
  The Next Web · 11:57 UTC
  Korbak, Balesni and Wang say they were never told what they leaked; Greenblatt calls the accusations false.
• Elvis Presley's rights holders sue Anthropic over 274 songs, seeking $150,000 a track, about $41 million
  Rolling Stone (Australia) · 00:44 UTC
  Authentic Brands Group's suit, filed Friday, says Claude trained on the songs and still reproduces them 'for a modern audience'.

The flaw under most websites (9)

• Retracted Palo Alto takes back its report of a government break-in
  Palo Alto Networks Unit 42 · 21:15 UTC
  'We are unable to confirm that initial access is associated with the specific PHP2Shell CVEs.' The only vendor account of a victim is withdrawn.
• Docomo restores J-Alert and evacuation alerts at 20:01 JST, earlier than planned, after PHP maintenance
  Keitai Watch · 11:34 UTC
  Area Mail suspended since 19:30 JST Oct 23; earthquake and tsunami warnings never stopped; full service back Sunday evening.
• KISA recommends a temporary fix for the third PHP bug, CVE-2026-115560, which the Oct 21 patches do not cover
  Topstar News · 06:12 UTC
  Guidance issued Oct 24; the agency also tells organizations to look for signs of intrusion, not only patch status.
• Nepal's government sites enter a fourth day offline; licences and the central bank still down, passports back
  Ratopati
  Shared GIWIMS platform awaits a central patch; Transport, Revenue Investigation, Land Management and Nepal Rastra Bank sites remain suspended.
• Korean banks widen PHP2Shell checks; regulators found security gaps at 130 of 555 financial firms inspected
  eToday (Korea) · 04:31 UTC
  Shinhan extends checks to unmonitored subsidiaries; Broadcom lists three affected products; the 555-firm inspection is not tied to PHP.
• Japan's PHP2Shell service suspensions pass 500, Assured counts; SoftBank's alert emails off until Oct 26
  FNN (Seoul), Tokyo correspondent · 02:48 UTC
  285 suspended services had no restart notice as of Oct 23; Orbis re-suspended its store, Bandai Namco still assessing.
• SoftBank gets quake and tsunami alerts back at 09:25 JST; evacuation alerts stay off until Oct 26
  Keitai Watch · 02:48 UTC
  Emergency-alert mail was halted at 06:47 JST Oct 23 for PHP maintenance; users pointed to Yahoo! Bousai Sokuho meanwhile.
• Samsung, SK, Hyundai, LG and HD Hyundai report no PHP2Shell breach; POSCO and GS patched in a rush, per Newsis
  Newsis · 02:05 UTC
  Big groups run little PHP; the worry, Newsis reports, is contractors and small firms on old versions that cannot be patched.
• Hacker News roundup: AI-paced PHP2Shell intrusion at a European government body; Bitsight counts nearly 37M PHP systems
  The Hacker News
  Unit 42's two-hour credential theft (retracted Oct 25), CYFIRMA's 216-host DPRK-suspected scan and Bitsight's exposure count; no victim named.

Everyday life (4)

• Third PHP flaw lands on Korean web shops with no security staff; KISA's workaround is to switch off upload tracking
  Financial News (Korea) · 20:30 UTC
  Shop operators say they can only wait on hosting providers; some had not heard of the flaws days after the Oct 22 alert.
• Dutch firms turn to free Chinese open-weight models; open models rise to about 70% of OpenRouter usage
  NOS · 18:51 UTC
  An Amsterdam law firm runs Alibaba, DeepSeek, Moonshot and Z.ai models for cost and control; China weighs licence rules.
• ITU deputy chief: Africa and Latin America hold 3.3% of data-centre capacity as AI's divide widens
  Al Jazeera · 17:26 UTC
  2.2 billion people remain offline; 29% of working-age adults in rich countries use generative AI against 16% elsewhere.
• Goldman slashes Korea target 25% as AI trade unwinds
  Investing.com
  Strategists say chip valuations imply a 65% drop in 2027 sector earnings; forward P/E near 4, unseen in 18 markets since 2006.

Deepfakes (6)

• Channel 12: a cyber firm working with Shin Bet dealt with about 80 Russia-linked fake accounts before the vote
  The Times of Israel · 18:52 UTC
  The profiles were primed to push 'violent and divisive discourse'; the report does not say whether AI was used or anything posted.
• Reported 43% of US voters say an AI ad fooled them
  GeekWire · 18:17 UTC
  Only 29% of AI ads carried labels in states that require them; new unlabeled examples in Oklahoma, Minnesota and New York.
• Judge's campaign pulls AI video of fake Black supporters
  Chron, via Yahoo News · 15:57 UTC
  Warped backgrounds and one voice for four speakers gave it away; the campaign has not said where the video came from.
• Netanyahu posts AI nuke hitting Tel Aviv
  The Times of Israel · 15:36 UTC
  Text by a right-wing activist warns of a 'second Holocaust'; Eisenkot, Golan, Bennett and Lapid respond.
• 26 court cases of alleged AI-faked evidence
  Reason (Volokh Conspiracy) · 12:01 UTC
  Damien Charlotin's 'AI Deepfakes in Court' logs fabricated texts, records and correspondence submitted as evidence.
• Al Jazeera: bots built against Iran now push Likud and both camps post deepfakes, two days before Israel votes
  Al Jazeera · 05:27 UTC
  Haaretz-reported bot repurposing; AI videos of Netanyahu and the opposition on X; Likud switches turnout apps after a privacy ruling.

Families and kids (2)

• Australia's education ministers will weigh a schools 'digital detox' next month after record-low PISA scores
  ABC News (Australia) · 18:00 UTC
  Minister Jason Clare cites Sweden's experience; Estonia's AI-tool embrace and Chinese AI pilots are the counter-examples.
• Reported AI school promised '2x learning'; most kids missed it
  Gizmodo · 09:00 UTC
  About half missed the math target and one in five the reading target at the AI-tutored chain; spokesperson declined interviews.

Jobs (1)

• Reported No AI payoff outside Big Tech, Apollo finds
  TechCrunch · 16:57 UTC
  Torsten Slok's note strips out the seven biggest tech firms and finds flat productivity measures for everyone else.

AI attacks (1)

• Reported One planted file hijacked OpenAI's coding AI 90% of the time
  BleepingComputer · 15:49 UTC
  Forever Security found the file on a customer's PC; Claude Code alone launched 1,300-plus sessions from Downloads folders.

Rules and regulators (2)

• AP: Kazakhstan joins both the US-led Pax Silica and China's 29-country AI body as nations hedge between blocs
  ABC News (AP) · 04:05 UTC
  Up to $10bn Nvidia data-centre deal on one side, record $2.8bn Chinese investment on the other; Washington says partners must choose.
• GAO: six of ten Pentagon AI teams found operational risks; the responsible-AI office has filled 3 of 10 posts
  NPR via OPB
  Four of ten projects were not aligned with the 2021 responsible-AI tenets; the Pentagon agreed with GAO but had not acted by Oct 22.

← Oct 24   Oct 26 →   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.