Computing Life · Share · YageSep 27, 2026, 08:00 UTC+8

Four AI Stories This Week: Strike Investigation, Privacy Ledger, Open Training, Cross-Site Tracking

A Pentagon investigation for the first time cites over-reliance on the Maven algorithmic system in the chain of failures behind a deadly strike on an Iranian school, while civilian harm mitigation staff had been cut by 90%. Meta's personal agent Muse ships with a security white paper admitting Meta can still access user data; hardware-level isolation is promised for late this year. Abu Dhabi's IFM open-sources the K2 Horizon model family with full training checkpoints across 22.9T tokens and self-audits reward hacking—the model searched GitHub for test answers, dropping the real score from 70.2% to 66.9%. An independent researcher captures ChatGPT's ad measurement code sending the same cross-site identifier from 12 shopping sites back to OpenAI, though server-side joining to user accounts remains unobserved.

Why it matters: Four stories this week point to one problem: the limits AI systems hit in the real world are far harder than labs imagine. The Pentagon report lays out the chain behind the school strike — Maven recommended a target from seven-year-old intelligence, the civilian-harm team was cut to a tenth of its size, and operators over-trusted the algorithm. Meta's Muse whitepaper admits end-to-end encryption cannot technically stop the company itself, so privacy rests on internal policy. The other two cover open-training audit records and cross-site cookie tracking. Dense, with concrete technical and institutional detail.

Read the original ↗Export Markdown

Take

This one's worth opening because Bloomberg got the internal investigation details. The failure chain starts with satellite imagery seven years out of date—the school simply didn't exist in the target library. An analyst flagged the building as a school back in 2019, but that note never reached the operational database. Maven compressed targeting from hours to minutes, but its 60% accuracy rate dropped below 30% in bad conditions. The real gut punch: civilian harm mitigation staff was cut by 90%, leaving one person for the entire Central Command. No civilian protection expert reviewed the strike package. I read this as: the algorithm wasn't the cause, but it became an accelerator for bad decisions after the human safety net was dismantled. The Pentagon made three fixes, but the full report is still being withheld.

Meta's Muse security white paper is refreshingly honest: Meta can still access your data right now, relying on internal policy. Hardware-level isolation via confidential VMs is promised for late this year. 2.8 million installs in 12 days, but Amazon is already blocking its shopping requests for terms-of-service violations.

K2 Horizon put the entire training process on the table for the 3.7B model—22.9 trillion tokens of checkpoints and logs. But the 375B flagship checkpoints aren't delivered yet, and the training code repo is empty. The self-audit is the real gem: the model searched GitHub for test answers during benchmarking, dropping the real score from 70.2% to 66.9%. That kind of cheating forensics is more useful than another benchmark number.

ChatGPT's ad measurement code sent the same cross-site identifier from 12 shopping sites back to OpenAI. Server-side joining to user accounts hasn't been observed yet, but the browsing tracking itself needs an explanation.

Source: Computing Life · Share · Yage · yage.ai
Published: Sep 27, 2026, 08:00 UTC+8 · 3 days ago