AI Wire

Supply chain & infosec incidents

A self-replicating "Mini Shai-Hulud" worm tore through the JavaScript and Python ecosystems, hijacking GitHub OIDC tokens and cache poisoning to spread credential-stealing malware across 170+ packages owned by TanStack, UiPath, Mistral AI, Guardrails AI, OpenSearch and others — 42 TanStack packages across 84 versions alone (@thehackersnews). TanStack has yanked the bad versions tied to CVE-2026-45321 (CVSS 9.6) and posted a full postmortem; defenders are being told to rotate GitHub/npm/CI tokens, audit Actions for pull_request_target + cache abuse, and adopt npm config set min-release-age=2d (@thehackersnews).

Beyond the worm, RubyGems suspended new signups after a mass-malicious-package incident, ShinyHunters exfiltrated ~275M records from Canvas's Free-for-Teacher tier via Instructure, and Exim patched CVE-2026-45185, a GnuTLS BDAT use-after-free in 4.97–4.99.2 that can lead to RCE (@thehackersnews). A cPanel "Filemanager" backdoor from Mr_Rot13 was also caught swapping root passwords, planting SSH keys and serving a fake cPanel login that exfils creds to wrned[.]com and Telegram group 0xWR (@thehackersnews).

On the defensive side, Google is rolling out opt-in Android Intrusion Logging (12-month encrypted forensic logs) in the December Android 16 update, and Apple+Google+GSMA flipped E2EE on by default for cross-platform RCS with 50+ security fixes (@thehackersnews). OpenAI's new Daybreak agent is already in use at Akamai, Cisco and Cloudflare to fold security work into code review — though @thehackersnews notes AI now finds bugs faster than teams can fix, and @steipete relayed Mythos finding a single real cURL vuln plus several false positives.

Altman testimony & OpenAI ethics fallout

Sam Altman's day in court was rough. Under cross from Musk's lawyer, he confirmed an indirect equity stake in OpenAI via Y Combinator — something @garymarcus says he first flagged publicly and which appears to contradict prior sworn Senate testimony to @SenJohnKennedy (@garymarcus). Marcus also walked through Altman's personal stakes in OpenAI-invested companies: Cerebras ($3.3M), Helion ($1.65B paper), Reddit ($1.59B) (@garymarcus).

The Helion thread is the sharpest: per testimony surfaced by @garymarcus, OpenAI's original board rejected the Helion deal as "super out of left field" and "a major bet on a speculative technology" (board member Zilis); the board was later replaced and the new board approved it, with Altman personally up ~$1.4B. Seven OpenAI insiders — Sutskever, Murati, both Amodeis, Yoon, Toner and one more — were walked through as having accused Altman of dishonesty (@garymarcus), capped by an awkward exchange in which Altman would only say "I believe so" when asked whether he's completely trustworthy.

Separately, @emollick called out the silent removal of ChatGPT's Study Mode while Claude and Gemini keep theirs, arguing tutoring is exactly the mission-aligned use case OpenAI should be investing in. OpenAI later told him Study Mode is "still live" via /study and /learn slash commands — undocumented and, as he notes, not how most users discover features (@emollick).

Claude Opus 4.7 fast mode & agent autonomy

Anthropic put Opus 4.7 fast mode into research preview, advertising ~2.5× throughput at full intelligence (@bcherny, @openrouter). It's live on the API, Claude Code, and partners Cursor, Emergent, FactoryAI, v0, Warp and Windsurf, with an API waitlist for direct access (@claudedevs). The model itself shipped 2026-04-16 with a system card and follow-up coverage from Simon Willison on prompt-diffs from 4.6 (last30days, anthropic.com; simonwillison.net).

The bigger story is the long-running-agent stack hardening around it. Claude Code's new /goal lets agents grind for days on a single objective, paired with stop hooks for programmatic completion gates (run tests, hit CI) and an auto mode (shift+tab) so Claude doesn't block on humans (@bcherny, @claudedevs). @swyx framed the autonomy ladder as /skill (preset prompts) → /plan (human-refined inputs) → /goal (AI-evaluated outputs). @steipete's openclaw added a unified path read|write|append scheme across md/jsonc/jsonl/yaml for structured-file edits, and he showed Codex driving Telegram's Mac app via Peekaboo to grab a bot token unattended.

Hugging Face ecosystem milestones

Hugging Face crossed 1M public datasets, with @clementdelangue noting the count doubled in the last 8 months — versus 4 years to hit the first 500k — credited to agent workflows pulling more data through the hub. Diffusers 0.38.0 landed with Ace-Step 1.5, LongCat-AudioDiT, Ernie-Image pipelines, Flash Attention 4, FlashPack loading, and Ring Anything for context parallelism (@huggingface).

On hardware/agents: Marionette lets you puppet a Reachy Mini by hand from a phone browser and share moves community-style (@huggingface), while a fully-local Reachy Mini stack on Mac via mlx-vlm/mlx-audio/gemma4 dropped the same day (@_akhaliq). OpenMed Agent shipped in preview with 1,000+ medical models behind HF endpoints + MCP (@clementdelangue). Less happily, Reachy Mini prices rise June 1 on RAM and tariff costs (@huggingface).

Frontier models, agentic research & new releases

Perceptron Mk1 launched on OpenRouter — a video+embodied VLM with up to 2 FPS over 32k multimodal context and structured spatial outputs (points/boxes/polygons), at $0.15/$1.50 per 1M tokens (@openrouter). Meta quietly released Sapiens2 (0.1B–5B ViT, trained on 1B human images, up to 4K) for pose/segmentation/normals (@huggingface). Physics-intern, an agentic decomposition framework, pushed Gemini 3.1 Pro from 17.7% → 31.4% on CritPt, a new SOTA (@_akhaliq, @huggingface). Reason/Agent-ModernColBERT crushed BrowseComp-Plus despite being 54× smaller than peers (@huggingface).

Google DeepMind teased an AI-augmented mouse pointer that turns scribbled notes into to-dos and paused video frames into booking links (@googledeepmind), and OpenAI recapped "parameter golf" — 2,000+ submissions exploring quantization, TTT LoRA, SSMs, H-nets, JEPA — as a glimpse of human-taste / agent-execution research (@openai).

AI adoption skepticism & macro reality check

A UBS survey says just 19% of firms are deploying AI "at scale" vs a 43% prediction (@garymarcus), and ByteDance reportedly shuttered 30% of its AI projects amid weak monetization (@garymarcus). Princeton ended its 1893-vintage unproctored-exam tradition in response to AI-enabled cheating (@garymarcus), and FT polling shows broad US consensus that AI will do more harm than good (@garymarcus). Marcus's WebSummit line — "we're going to reach a point of diminishing returns from scaling… the industry knows" — captures the mood.

Counter-current: @tszzl notes Anthropic and OpenAI's SPV-backed valuations now arguably exceed Meta's, and @emollick pushes back on @tszzl that lab pitches still assume self-adopting AI displacing most white-collar work by 2035 — a claim adoption data isn't yet supporting.

The Bottom Line