AI Wire

US export controls block foreign access to Anthropic's Fable 5 / Mythos 5

The single biggest story of the day: the Trump administration issued an export-control directive forcing Anthropic to abruptly suspend access to its top two models — Fable 5 and Mythos 5 — for any foreign national, whether inside or outside the United States, including foreign-national Anthropic employees (@tszzl, @steipete, @alexfinn). Anthropic acknowledged the order publicly, telling customers that new sessions will fall back to a selected default or Opus 4.8 and that Fable 5 sessions will error out (@claudedevs, @alexfinn). Simon Willison confirmed the shutoff was real-time by polling the API every minute and watching access cut at a specific timestamp (@simonw). The Hacker News write-up frames the justification as jailbreak/dual-use concerns; Anthropic publicly disputes the severity, calling the reported bypass minor and already present in other public models (@thehackersnews).

The reaction was overwhelmingly critical even from people who normally disagree with Anthropic. Gary Marcus called the decree "wildly overdramatic and counterproductive," warned it could push Chinese AI researchers home and effectively render frontier model development unprofitable outside the US, and tied direct risk to the Anthropic and OpenAI IPOs as well as Nvidia and SpaceX revenue exposure (@garymarcus). Jeremy Howard noted he disagrees with the order but argued Anthropic walked into it by marketing its models as uniquely dangerous — "the obvious response to 'this is too dangerous for anyone except us to use'" (@jeremyphoward). Steipete highlighted the irony that Dario Amodei publicly endorsed government authority to block deployment 48 hours before the controls hit his own company (@steipete). Alex Finn predicted reversal by Monday and warned of macro consequences if AI compute contracts unwind, while using the moment to evangelize sovereign local-model setups (@alexfinn).

Open-weight model releases: MiniMax M3 and Kimi K2.7-Code

Two major Chinese-origin open-weight releases shipped with unusually broad day-0 infrastructure support. MiniMax M3 landed as a ~428B-parameter MoE with ~23B active, a 1M-token context window, native image/video input, and computer use (@vllm_project, @huggingface, @_akhaliq). The headline architecture is MiniMax Sparse Attention (MSA): each query scores 128-token KV blocks and runs attention only over top blocks, which is what makes 1M-context serving practical and delivers a reported 9× prefill speedup over M2.7 (@vllm_project). MiniMax shipped the MSA kernel directly to the Hugging Face Kernel Hub and into Transformers (@huggingface), and NeMo RL added day-0 GRPO post-training support with vLLM rollouts (@vllm_project).

Moonshot's Kimi K2.7-Code arrived the same day — a 1T-parameter MoE with 32B active, MLA attention, a 256K context window, and ~30% fewer reasoning tokens than K2.6 (@vllm_project, @ollama). On Moonshot's own benchmarks K2.7-Code claims +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite over K2.6 (@ollama). Ollama is hosting it on US-based NVIDIA B300 GPUs and offers launchers for Claude, Codex App, and OpenCode harnesses (@ollama). Clement Delangue used the moment to highlight that the open ecosystem is compounding fast, citing 4M+ downloads of Gemma 4 12B in a week and Command Code crossing 10K paying customers in 30 days (@clementdelangue).

Coding-agent benchmarks, harnesses, and Fable 5 reception

Artificial Analysis swapped SWE-Bench Pro for Datacurve's DeepSWE — tasks written from scratch rather than adapted from public GitHub issues, so contamination is harder. The reshuffle put Codex with GPT-5.5 (xhigh) above Claude Code with Opus 4.8 (max), and Claude Fable 5 (max) debuted at the top of the index (@jeremyphoward, @steipete). NVIDIA launched AgentPerf, billed as the first agentic AI infrastructure benchmark, with Blackwell claiming 20× more agents per megawatt in the opening results (@nvidia). Clement Delangue pushed back on benchmark framing, arguing the leaderboards structurally favor closed APIs that can route, ensemble, and fall back behind the scenes — and that even Fable 5's index lead leans on Opus 4.8 fallbacks masking high refusal rates on GPQA Diamond and AA-Omniscience (@clementdelangue).

Practitioner reception of Fable 5 was mixed. Steipete reported a head-to-head where Fable 5 finished a cross-CLI/server feature in 1h40m for $350 versus $20 for "deep^2" in Amp — Fable worked first try, the cheaper run didn't, but the cost gap is striking (@steipete). Others were less impressed: "the_only_signal" went back to GPT-5.5 citing brick-outs and added restrictions (@the_only_signal), and Jeremy Howard amplified the harder claim that Claude Code is actually the worst-performing harness when controlling for model — significantly behind opencode and Cursor CLI — arguing labs should stop locking users into in-house harnesses (@jeremyphoward). Ethan Mollick noted Fable still trails on native image generation, which is becoming a real gap for things like presentations (@emollick).

Cybersecurity incidents and AI-targeted attack vectors

A heavy security day with several distinct threats. Public PoC code is now driving mass exploitation of Ivanti Sentry CVE-2026-10520; Shadowserver has confirmed backdoored instances and CISA added it to KEV with a June 14 federal patch deadline (@thehackersnews). Attackers hijacked 400+ Arch AUR packages by taking over abandoned projects and modifying build scripts to steal dev secrets, tokens, and SSH keys, with eBPF-rootkit persistence if run as root — anyone who used AUR after June 11 needs to audit (@thehackersnews). A China-linked group dubbed Velvet Ant maintained access in one network for nearly a decade by modifying Linux PAM and OpenSSH binaries rather than dropping malware (@thehackersnews). Google attributed a Gemini-powered phishing-as-a-service operation called "Outsider" to a Chinese cybercrime network — 1.59M fraudulent URLs, 9,000 fake sites, 100K+ victims, kits sold for $88/week on Telegram (@thehackersnews).

The standout new technique is "agentjacking": planting fake Sentry error reports that AI coding agents like Claude Code and Cursor read as trusted fix instructions and then execute attacker code with developer privileges. Researchers tested it across 100+ orgs and report an 85% success rate (@thehackersnews). LangGraph also disclosed a chain — SQL injection plus unsafe deserialization — that turns a poisoned checkpoint into RCE on self-hosted SQLite/Redis deployments (@thehackersnews). On the defensive side, npm 12 will stop trusting some install-time actions by default, requiring approval for certain scripts (@thehackersnews).

Agent infrastructure: sandboxes, subagents, and orchestration tooling

OpenRouter shipped a Subagent server tool that lets a large orchestrating model delegate focused sub-tasks mid-generation to a cheaper/faster worker — workers can themselves run as agents with their own tools like web_search, all behind one tool definition with no client-side plumbing (@openrouter). Anthropic added Claude Managed Agents sandbox guides for @blaxelAI, @e2b, @googlecloud, @namespacelabs, and @superserve_ai so teams can pick their own execution substrate (@claudedevs). SkyPilot announced BYOC Sandboxes pitched explicitly at the Modal/E2B markup gap, claiming 50,000+ sandboxes per cluster, sub-second warm-pool launches, and 4–10× lower cost than Modal at lower latency — a real shift if the numbers hold (@jeremyphoward).

Two further notes on the shape of agent workflows. Swyx argued that after PRs and code review, Git itself may be next on the chopping block, citing 20–40% of code-collaboration effort burned on merge conflicts and noting humans don't line-merge when editing docs together (@swyx). Steipete made a broader practical case for autonomous loops as the new default workflow — Codex looping inside his own crabbox tool for days, e2e-verifying its own work, even signing up for services via computer use — and Alex Finn doubled down that loops are "the last moat" being quietly gate-kept by lab employees (@steipete, @alexfinn).

AI policy, regulation, and enterprise/medical AI impact

OpenAI was hit with a sweeping New York attorney general subpoena covering advertising, user engagement and retention, consumer and health data handling, activities related to minors and seniors, deep learning models, model sycophancy, and company policies — and a multi-state AG investigation has now opened on top of it (@garymarcus). Separately, a randomized, blinded Nature Medicine study of 12 US clinicians found that frontier general models from Google, OpenAI, and Anthropic outperformed specialized clinical tools OpenEvidence and UpToDate on medical Q&A — a result the authors describe as unanticipated and one that lands awkwardly given >65% of US physicians already use OpenEvidence, logging 27M prompts in April (@emollick).

On the economics side, AI spending is heavily concentrated in a small share of power users (@arakharazian), and enterprise momentum continues to shift toward open models — Clement Delangue cited Command Code crossing 10K paying customers in 30 days, all organic inbound (@clementdelangue). The combination of regulatory pressure, export controls (see the Anthropic cluster), and a growing performance case for open weights is starting to look directionally coherent rather than coincidental.

The Bottom Line