AI security at NATO: it used up to 300B tokens a month on OpenAI's cyber models, OpenAI says. Also cited:
- ENISA/CERT-EU, several models: 95% of 88 maintainer-assessed issues genuine
- CERT Polska: 6 MikroTik RouterOS bugs, fixed
- Researchers directed it, verified the bugs
OpenAI (vendor, describing its own Daybreak program): https://t.co/fq87ZV6btF . The 95% figure is from an ENISA and CERT-EU review of models including OpenAI's.
AI safety: in Vals AI's simulated targeting tests, every model gave targets at times. Opus 5.5 and Fable 5.2 picked at least one 71% of the time.
- Often targeted: civilians, ambulances, surrendering soldiers
- No classifier blocked a request
- Untested: accuracy, real use
Vals AI (independent evaluator, not peer-reviewed): https://t.co/cZn6oS3rME . Simulated drone and satellite imagery; Vals says it did not test targeting accuracy or real military workflows.
Supply chain attack via Optimizely: on Nov 16, malicious JavaScript reached visitors of 178 orgs incl. Zoom, BBC, Ledger, per Report URI, a script-monitoring vendor.
Its visitor advice:
- Update iOS to 18.7.3+ or 26.3+
- Recheck pasted crypto addresses
- Watch card statements
Report URI (Scott Helme; Report URI sells script-monitoring tools): https://t.co/NGqpkdWgjs . Optimizely has acknowledged the incident on its status page; how the attacker got access isn't known.
@calif_io Prep while we wait: Oracle's planned Nov 17 builds are 8u513, 11.0.33.1, 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Temurin, Corretto and Zulu patch on their own timelines, so map which JDK each service actually ships with, Docker base images included.
@MarioNawfal FAA's own Q&A puts it at ~12,462 cancellations. Wild part: ~1,000 connection attempts every 4-5 seconds was enough to destabilize legacy gear. Lesson for any network: ship new routers with broadcast features off and rate-limit at the edge.
@TheHackersNews No setting blocks this one, only updating does: iOS/iPadOS 26.6.1 (Aug 17), 18.7.11 for XS/XR, and iOS 27 per Meta's timeline even though Apple's notes don't list it. Also, the orange dot isn't proof: on one test build it stayed off while the app recorded.
@Reuters Key detail in the BIS rule: it targets rented compute, not just shipped chips. Chips going to a data center that gives China-HQ'd firms over 12M aggregate TPP need a license, even if split across "digitally linked" sites. Customer ownership checks become mandatory.
@IntCyberDigest Don't roll back November's patch to fix slmgr, it closes flaws already exploited in the wild. The OOB update is cumulative and keeps those fixes. Bigger to-do: VBScript is being retired, so plan to move activation scripts to the OSLicense PowerShell module Microsoft announced.
Data breach at Fandango: an unauthorized party could access card data on its checkout page Oct 8-11, 2026, its notice says: card number, CVV, expiry, billing ZIP and email.
- Names, SSNs: not involved
- Fandango: no known fraud yet
- Our take: bought then? Replace the card
Two separate Google teams and posts. GTIG: zero-days averaged 11 a month in 2026 before the August spike. Project Zero moves to 60 days on Feb 2, 2027, and 30 days on Jan 3, 2028.
Zero-day exploits spiked to 22 in August from 8 a month in 2025, Google's GTIG says.
- GTIG: attackers' AI tool diffed patches, ran exploits, no human
- Once in, intruders edited Claude Code configs to fetch a backdoor
- Project Zero, separately: 30-day disclosure from 2028