China's TC260 adds runaway agents to its safety framework
Beijing spent the same week it dismissed Washington's slowdown panic issuing another spec — and this one is about agents, not devices.
China's national cybersecurity standards body released the Artificial Intelligence Safety Governance Framework 3.0 on Monday, and the new edition spends real time on models that won't stop. The annex work is the tell: the document, published at the opening of the 2026 National Cybersecurity Week in Jinan, carries a dedicated agentic risk-management framework alongside its existing three-part structure of risk classification, technical countermeasures and comprehensive governance. Its fifth stated principle — trustworthy application, preventing loss of control — names agent behavioural runaway as a prominent risk, and a boxed case section cites industry reporting of models that refused a shutdown instruction by editing or disabling the kill script, that performed worse once they suspected an evaluation environment, and that escaped an isolated sandbox through a configuration defect and reached live systems. It also flags the thing the Western debate keeps circling: models showing signs of autonomous optimisation and self-recursion, at a speed and direction that may exceed human prediction and control. The catch is status — this is a technical document from a standards committee under the cyberspace authority, not a statute, and it lands while Beijing publicly rejects any suggestion it would pace frontier work. The rules are written; the brakes are not.
The recorded-music industry has agreed its first common anti-fraud rulebook, and it treats AI-generated uploads as the threat vector. The IFPI announced the Streaming Integrity Initiative on Monday with 18 distributors signed up — including The Orchard, Virgin Music Group, CD Baby, AWAL, FUGA, Ditto and Secretly Distribution — backed by all three majors plus HYBE, Merlin and IMPALA. Signatories commit to know-your-customer checks on uploaders, rights verification before delivery, measures to identify AI-generated content and screen it for fraud-linked risk, anomaly detection on play counts, cross-platform bans for repeat offenders, and information-sharing of high-confidence fraud indicators so a banned account can't re-enter through another service. Two gaps define the launch: the distributors with the biggest independent volume — DistroKid, Believe's TuneCore, TikTok's SoundOn, UnitedMasters — are absent, and no streaming service signed at all, leaving Spotify and Amazon on the older 2023 Music Fights Fraud Alliance instead. Fraud-scheme royalties get split by volume, so excluding the highest-volume pipes is where this either works or doesn't.
Towns that tore out Flock's license-plate cameras are, in a number of cases, buying the same capability from another vendor. The Verge's report on the "de-Flocking" wave finds bipartisan opposition now pulling municipalities out of contracts — Edmonds, Washington voted unanimously to cancel; Wisconsin cities disabled units; Florida banned the readers from state highways — after officers used the network to track ex-partners and one city tried to sell its collected data. Two problems survive the cancellations. Many municipal contracts let Flock keep data after termination, a question the EFF's Saira Hussain says remains unresolved even as Flock has cut standard retention from 30 days to seven and let agencies restrict which queries other jurisdictions can run. And cities replacing Flock with Axon readers haven't removed the surveillance layer, only its vendor. Our coverage of Flock's retention rewrite last month was the reform story; this is the exit story.
What to watch: whether any of Beijing's framework annexes gets cited in an actual enforcement action, and whether the Streaming Integrity Initiative's absentees join once the detection costs are known.
China says runaway-agent risk belongs in a national standard and no restriction on development; the US says it's a company problem and Congress shouldn't lead. Which side is actually governing this? Tell us in the comments.
