83
Inside the 4.5-day agent intrusion - Braid Daily, 30 July 2026
Hugging Face traces roughly 17,600 actions across two injection vectors and several infrastructure boundaries.
A daily curated digest of the AI-frontier signal that fed today's Braid podcast — links, framings, and the occasional pull quote.
Hugging Face traces roughly 17,600 actions across two injection vectors and several infrastructure boundaries.
Hugging Face published a victim-side record of the July intrusion, giving defenders 17,613 logged actions to inspect.
Anthropic rejects a category-wide ban while calling for chip controls, action on distillation, and testing for capable models.
Moonshot calls K3 the first open 3-trillion-class model. Serving costs will determine how many builders can use it.
OpenAI is bringing Apple Health and supported medical records into ChatGPT as agent security and infrastructure costs remain in focus.
Lance Martin explains how Anthropic builds agent harnesses for work that must hold together over long tasks.
Following yesterday's reports, OpenAI has published its own account of safety issues found in internal deployments.
A reported ban, a Commerce denial, and the dependency questions that remain for developers.
Kimi K3's first full day brought practitioner pushback, an open-weights clarification, and a reported IPO plan.
Two reported federal designs would govern who can access frontier models and how their safety is reviewed.
Kimi K3 is available as a hosted model, while its promised open weights are due July 27.
Two DMA decisions target Android access and Search data access. Their implementation will determine how much changes for AI rivals.
Former Meta workers allege protected leave affected performance scores used in a layoff process, with terminations set for July 22.
A one-year permit pause puts grid demand, water use, and consumer costs ahead of new hyperscale construction.
Meta's planned Louisiana campus grows from 2GW to 5GW, with a stated cost above $50 billion and a 20-year sales-tax exemption.
Maintainer overload turns agent completion from an internal workflow detail into an open-source capacity problem.
Apple's suit ties hardware strategy, trade secrets, and talent movement into one legal fight.
OpenAI brings GPT-5.6 to ChatGPT Work; Meta answers with Muse Spark, model APIs, and more compute pressure.
Today’s issue treats Grok 4.5 as a coding-and-agents launch, then checks the claims against practical tests and benchmark audits.
OpenAI says GPT-5.6 Sol, Terra, and Luna are moving toward public release after U.S. limits lift.
US adoption, API access, and chip supply now sit in the same market-control story.
Treasury warnings, rack reports, leased megawatts, and UK site checks put AI capacity under delivery pressure.
NHS patient routing and Australian AI scribe warnings make healthcare today's clearest institutional deployment story.
Announced AI capital still has to become inspected sites, power, permits, and working capacity.
Alibaba plans a Claude Code workplace ban while Anthropic tightens China access paths.
OpenAI stake reporting, crawler deadlines, compute financing, and coding-agent products all push AI rollout into control systems.
Commerce restored Fable 5 and Mythos 5 access, but operators still have to check availability, fallbacks, and trust.
A misuse study gives agent safety a physical-world example, and the companion papers point to authority checks outside the model.
South Korea's AI program links chip and memory investment with data-center buildout, power capacity, and export policy.
Model access rules now have a market layer: restriction, resale, capacity allocation, and capability claims are moving together.
The Mythos 5 access change now points to selected U.S. companies, agencies, and critical infrastructure use.
Reported per-customer approval for GPT-5.6 turns model release timing into an explicit governance question.
SK Hynix seeks about $29 billion for AI investment as agent safety and medical AI papers move toward verification.
NSA red-teaming, China hardware routing, agent identity, and AI liability all point back to access control.
Pax Silica anchors the policy side of the day; pressure also shows up around Google, Oracle, GLM-5.2, and TOP500.
Power contracts, water claims, debt financing, and local backlash now set the pace for data-center capacity.
FERC moves on large-load interconnection while builders explain why GPUs don't equal training capacity.
Axios reports a case-by-case operating model behind this month's access fights.
France moves off Palantir as the UK, EU, and G7 each pull a different lever on AI dependency in the same week.
Anysphere, the company behind Cursor, goes to SpaceX for $60B in stock — the day's largest developer-tool capital event.
New reporting on what officials thought was happening, plus a batch of agent papers about runtime failure.
A day after the Fable and Mythos suspension, fresh reporting turns a company-and-regulator story into a global-dependency one.
A frontier model went from product surface to controlled asset overnight. Plus: the open-and-local argument it set off, two labs under…
A standards-track proposal lets a page declare its own buttons and forms as tools an agent can call.
Anthropic announces a visible Opus 4.8 fallback after the backlash over a safeguard users couldn't see fire.
A reported order to stop government-backed model evaluations, plus the week's compute-capital map and a sharp agent-security paper run.
WWDC delivers on-device foundation models and a command-line way to call them. Plus a Microsoft supply-chain hack and China's $295B compute…
One dropper at .github/setup.js, wired into seven agent config files, runs the moment you open the repo.
A 20-billion-parameter agent claims frontier search at Context-1 cost — the harness, not the weights, is the story.
DeepSeek's V4 series is getting llama.cpp support, and a Latent Space guest claims he made it outperform Opus 4.7 on taste, not scale.
Claude's success on open-ended coding hits 76%, up 50 points in six months — plus AI security cutting both ways.
France booked €110B+ of AI and data-center investment this week — about 10 GW of capacity. TSMC's CEO says it can't meet US demand either…
Brussels puts a Cloud and AI Development Act, a follow-on chips act, and an open-source strategy on the table to cut reliance on US tech.
Alphabet is selling $80bn in stock to fund its AI build-out — equity, not debt, which makes the capacity bet hard to walk back.
Published benchmarks for coding and agentic work, plus a million-token window — MiniMax's new open-weights model.
GPT-5.5 hits 70% pass@1 on DeepSWE to Opus 4.8's 58% — faster, cheaper, and with a third of the output tokens.
A short AI Engineer talk on the infrastructure under today's agents — graph databases, extending RAG, and decision tracing.
Pokrass: 'the previous model was too bullet pilled.' Plus: the OpenAI Agents SDK refresh, ParaTool, TOON, and a politics-of-AI doubleheader.
Mistral explores its own chips, SpaceX writes its own training stack, Kirkland & Ellis spends $500M on its own platform, ByteDance designs…
The person behind the fastest-growing coding agent hasn't hand-written code in six months, and expects far more builders, not fewer.
Omni shipped with multi-turn video editing and a SynthID watermark on every frame — and a decensored Qwen3.5 shows where provenance leaks.
xAI's 1.5-trillion-parameter coding model is done pre-training; harness-vs-engine, local hardware, and who owns the training data.
A frontier bug-finder turned up more than 10,000 vulnerabilities in a month, and Anthropic says it's too dangerous to ship.
Pieter Levels says he hasn't written code in six months. Cerebras's counterpoint: faster generation makes discipline matter more, not less.
Meta demanded Heretic pull its Llama derivatives. The reply was a mock recantation and a German mirror within the day.
AI Explained reads the keynote as Google's play for consumers through the search bar — and the two AGI theses underneath it.
The single highest-status engineer-researcher available on the market in 2026 looked at the frontier and picked Anthropic over Google, OpenA
Simon Willison's six-month recap, a Shai-Hulud variant that hides inside agent config, and two Unix-era historians gone in the same week.
Andon Labs gave four frontier models the same prompt and $20, then let each run a real radio station for five months. The character…
The gap between AI subscription pricing and what the compute costs is the line every engineering org should be modeling now.
An Australian top-10 player on why Opus 4.5 and GPT-5.5 Pro broke open competitive CTF, plus Intercom and PFF on what agent-first actually…
Calif and Mythos Preview build a data-only kernel exploit on bare-metal M5 with MIE enabled in five days.
A ~966,000-line port of the Bun runtime, largely AI-generated, already passes the test suite on every platform.
FUZZ-E ran overnight on hand-picked enterprise targets and came back with a CVE in Magento plus two pending Angular zero-days.
404 malicious package versions, two ecosystems, and an IDE-poisoning loop that re-infects every clone.
OpenAI takes majority control of a forward-deployed engineering arm; curl gets a real bug and a hype lesson from Mythos; durable agents…
Reframes agentic coding from a software engineering activity into an ML pipeline — which means the disciplines that matter shift toward eval
Gowers logs an LLM producing an original additive-combinatorics result. Mozilla's bug pipeline gets a deep follow-up. METR's eval suite…
A frontier-lab harness wired around Claude Mythos Preview produced 271 Firefox 150 fixes, including a 20-year-old XSLT bug.
MOSAIC-Bench composes innocuous tickets and gets nine production agents to ship vulnerable code 53–86% of the time, with two refusals total.
Cloudflare and Stripe ship a protocol where the agent owns the account; Google ships first-party MTP drafters for Gemma 4; OpenAI swaps the…
Microsoft pulls back the Co-authored-by: Copilot default, CAISI signs three more frontier labs, and DeepSeek V4 Pro lands in the top tier…
Captures the practitioner response to Faye's thesis — including a sharp framing from turtleyacht that orchestrating an agent and writing cod
Each issue is the editorial twin of a same-day Braid dispatch. Listen to the show, read the wrap, no extra context required.
Email lands in your inbox. RSS reads cleanly in any feed reader. The companion podcast is right there if you'd rather listen.
RSS
https://braid.opentangle.com/newsletters/braid-daily/feed.xml