← All news

AI News — June 29, 2026

#anthropic #open-weights #safety #benchmark

A wave of independent benchmarks this week put Zhipu's freely downloadable, MIT-licensed GLM-5.2 at or near restricted US frontier models on cybersecurity work — the exact capability the June 12 Fable 5 / Mythos 5 export ban was meant to contain — with CNBC clocking it within a point of Opus 4.8 on agentic tasks at roughly a fifth of the cost, the first concrete sign that API-level export controls can't hold a capability once an open-weight model reaches it.

Model releases

  • [2026-06-28] Zhipu (Z.ai) — a run of independent benchmarks this week put the freely downloadable, MIT-licensed GLM-5.2 at or near restricted US frontier models on cybersecurity work — the exact capability the June 12 Fable 5 / Mythos 5 export ban was meant to contain. Two security firms drove the coverage: Semgrep reported GLM-5.2 scoring 39% F1 on IDOR vulnerability detection, ahead of Claude Code’s 32%, and Graphistry’s agentic-cyber evaluation found it topping open-weights rivals and tying the closed proprietary models; CNBC separately clocked GLM-5.2 within a point of Opus 4.8 on a watched agentic benchmark at roughly a fifth of the cost. The Mythos comparison is inferential, not head-to-head — Mythos has been export-restricted since June 12, so the evaluations lean on Opus 4.8 and Claude Code as proxies, and some developers caution the parity claim rests on a narrow test set. It matters because it is the first concrete evidence that API-level export controls can’t contain a capability once an open-weight model reaches it: GLM-5.2 is downloadable by anyone, carries “no regional limits,” and can’t be revoked — which is also why CNBC reports enterprises are increasingly treating it as the safer bet while US frontier access stays gated. (CNBC, Axios)

    What changes for a practitioner here isn’t the leaderboard position but the deployment model: GLM-5.2’s weights can be pulled and run inside an air-gapped or compliance-bound environment where shipping code to a hosted US frontier API was never an option to begin with.

    For Security Engineers: If you triage code that can’t leave your network — regulated, customer-owned, or classified — GLM-5.2 is now a self-hostable candidate for that work; Semgrep’s run had it flagging IDOR flaws at 39% F1, ahead of Claude Code’s 32%, so it’s worth scoring against your own vulnerability corpus before assuming a gated frontier API is the only path.

Watch list

  • Fable 5 general release — day 17 of the blackout and the consumer-grade model is still dark. Axios’s June 27 scoop put restoration “as soon as this coming week” with Anthropic confident “in the coming days,” but the Pentagon and NSA still have to clear it, and the July 8 consumer-terms update enumerating identity-verification data remains the most-cited mechanism for a US-users-first return. GLM-5.2 matching the restricted capability sharpens the question: the longer Fable 5 stays gated, the more an open-weight alternative fills the gap. Watch for an Anthropic announcement and whether access returns gated by verified identity or as a plain public unlock. (Axios, prior coverage)
  • GPT-5.6 general availability — Sol, Terra, and Luna remain a ~20-partner government-coordinated limited preview; OpenAI says GA is “in the coming weeks.” Watch for the GA date, whether the vetted-partner gate persists at GA, and published third-party benchmarks beyond the internal CTF cyber scores. (prior coverage)
  • Gemini 3.5 Pro (GA) — month-end arrives tomorrow with the model still a limited Vertex AI enterprise preview: no model card, no public benchmarks, and no pricing for the promised 2M-token context and Deep Think mode. With Pichai’s Google I/O “give us until next month” now expired, watch for whether a rate card lands or the GA target quietly slides into Q3. (prior coverage)
  • Alibaba / Qwen distillation amendment — the Hagerty–Kim amendment to must-pass defense legislation (which would blacklist or sanction Chinese firms found improperly accessing US AI-model output) was floated for as early as June 24; that date has passed without a confirmed filing. Watch for whether it has been formally entered into the bill’s text and Alibaba’s still-absent response. (prior coverage)
  • MCP spec finalization (July 28) — the 2026-07-28 release candidate (stateless core, Extensions framework, Tasks, MCP Apps, authorization hardening, formal deprecation policy) is in its validation window, with the Ruby/TypeScript/Python SDKs being updated against it. Watch for the rest of the SDK support landing — and any breaking issues surfaced in Tasks, MCP Apps, or the new MCP-specific HTTP headers — before the cutover. (official, prior coverage)