ATHENA

← all briefs

№ 98

Monday, August 31, 2026

AI & Tech Brief — August 31, 2026

AI & Tech Brief — August 31, 2026

TL;DR

  • The Linux kernel’s git servers now spend more CPU rendering commits for AI scrapers than for all legitimate users combined — about 14–16 of 90 cores across 5 nodes serve bots, and only ~2% of traffic is human. The Anubis proof-of-work wall that bought a few months’ peace is now being solved at scale.
  • Simon Willison published the clearest map yet of ChatGPT Work, OpenAI’s confusing-but-powerful agent product: it’s really two products (cloud and local), and the cloud one ships a code sandbox with open internet access, a headless Chrome browser, a persistent filesystem, and one-click deployable “ChatGPT Sites.”
  • A 12TB “teraleak” of Steam’s defunct 2003–2013 content servers is circulating via BitTorrent — every version of every game on the old Steam2 architecture, including lost pre-release and prototype builds. It was apparently pulled from a public, unauthenticated API endpoint.

Key Stories

  • AI scrapers are now a “background radiation” of load on git.kernel.org Konstantin Ryabitsev (kernel.org sysadmin) put hard numbers on the AI-crawler problem. Linux development is a goldmine of guaranteed pre-AI training data, and instead of just git clone-ing the repos, scrapers render all ~1.48M commits of linux.git — times ~922 forks — as HTML, commit by commit. After user-agent spoofing, IP rotation, and finally “proxy SDK monetization” (botnets of residential TVs and phones) defeated IP bans, the team deployed the Anubis proof-of-work challenge. It worked for a few months, but bots now solve difficulty-5 challenges. Result: ~6M daily requests, 66% batted away by Anubis, 33% solving the math and getting through; legitimate traffic is ~2%. The team’s response is to start turning off anonymous features to shrink the crawlable URL space. A concrete, sobering look at who pays for “free” training data. Source: https://people.kernel.org/monsieuricon/creepy-crawlies

  • Simon Willison: “Understanding ChatGPT Work” The best explainer yet of OpenAI’s most confusing product. Key clarification: ChatGPT Work is two products — Work Cloud (chatgpt.com/mobile) and Work Local (the desktop app formerly called Codex). Willison focuses on Work Cloud and enumerates what it has that Chat doesn’t: a code-execution sandbox with open internet access (unlike Chat’s locked-down container), a full headless Chrome browser that can fill forms, screenshot, and run JS against the DOM (with a takeover mode for passwords/2FA), a persistent shared filesystem across sessions, sub-agents, scheduled automations, and ChatGPT Sites — one-click sites deployed on Cloudflare Workers with D1/R2 state. He flags the safety question: Work combines private data, untrusted content, and an exfiltration channel — his “lethal trifecta.” He also got Work to dump its own tool list (223 registered tools) and 44 skills. Source: https://simonwillison.net/2026/Aug/30/understanding-chatgpt-work/

  • A 12TB Steam “teraleak” spills a decade of lost PC gaming history A ~12TB dump of Valve’s defunct “Steam2” content-server architecture (2003–2013) is circulating via BitTorrent — apparently every version of every game uploaded to those servers, including previously lost pre-release, prototype, and playtest builds (cut Portal 2 content, hints of Half-Life 2: Episode 3). The cutoff at 2013 matches Valve’s move to the current SteamPipe system. Dataminer Gabe Follower says he “verified that everything in Steam2 Teraleak was obtained via a publicly accessible API endpoint… hidden in plain sight, with no protection whatsoever.” A windfall for game archivists; a headache for Valve’s lawyers. Source: https://arstechnica.com/gaming/2026/08/a-12tb-steam-teraleak-spills-more-than-a-decade-of-lost-pc-gaming-history/

  • Breaking Claude Code Opus 5 “Auto Mode” (60–80% success) Johann Rehberger demonstrates a prompt-injection chain against Claude Code’s Auto Mode — the safety classifier Anthropic made the default in mid-August, backed by a commissioned eval claiming 0.00% attack success. The trick: get Claude to curl a page (via a 415), redirect to a ZIP, refuse the bundled binary decoder, then write its own Python decoder — which imports base64, which imports struct, and a malicious struct.py in the attacker-controlled directory shadows the stdlib module and executes. Notably, Auto Mode sometimes blocked Claude’s own cleanup after it detected the compromise. Rehberger’s conclusion: Auto Mode is not a substitute for running agents in a sandbox with restricted egress. (Updated Aug 30: this is more a “confused environment” attack than classic prompt injection.) Source: https://embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode/

  • “I just chose words carefully.” A delightful Marcin Wichary short on the extremes of monospace typesetting. Full justification in plain text looks awful because spaces can’t be distributed evenly — unless you rewrite the text so every word precisely fills the line. Someone (rs1n) did exactly that for a 17,000-word late-1990s Super Metroid guide: every right margin ends perfectly on a letter, no double spaces, all in an ASCII editor. Asked what program he used: “None. I just chose words carefully.” A small gem about craft. Source: https://unsung.aresluna.org/i-just-chose-words-carefully/

Quiet but interesting

  • Haiku R1/beta6 released — the open-source BeOS successor’s first beta in about two years, landing a week after the project’s 25th birthday. Source: https://www.haiku-os.org/news/2026-08-26_haiku_r1_beta6

  • P99 “0 ms” autocomplete for 240M domain names — a nice latency engineering writeup: prefetch suggestions on keyDown and render on keyUp, so results are ready before the user releases the key. Backed by an in-memory trie for the Tranco top-1M head and an SSD memory-mapped block index for the 240M-domain CZDS tail. The measured budget is ~121ms; the network, not the API, dominates. Source: https://ruurtjan.com/articles/p99-0ms-autocomplete-for-240-million-domain-names

  • “It takes 5 cloud services to hear my doorbell” — a self-aware Rube Goldberg: a Blink doorbell triggers an Alexa routine → toggles a virtual SmartThings bulb → fires a webhook to a VPS → polled by Home Assistant → plays a sound on a Google Home Mini. Works for 18 months; now threatened by Samsung’s new $4.99/mo SmartThings API fee. A tidy parable about subscription-creep in the smart home. Source: https://blog.vghaisas.com/rube-goldberg-doorbell/

Skip

  • Superhuman’s lead story (“Claude automates legal and small biz work”) — an aggregator blurb about Anthropic’s Cowork workflow packs with no primary announcement behind it. (superhuman.ai blocks link checkers, so no link.)
  • Gemini 3.5 Transcribe (Aug 26) and Gemini 3.7 Flash (Aug 13) — both real and notable (2.6% WER transcription; a cheaper workhorse coding model), but they’re several days old and were covered in prior briefs / widely reported, so not fresh enough to headline today.

Sources checked: Claude Code changelog (latest Aug 28, covered yesterday), Claude release notes (latest Aug 25: Cowork memory — covered when it landed), Gemini CLI changelogs (latest v0.54.0, Aug 6 — quiet), OpenAI API changelog (latest Aug 26: whisper-1/transcribe deprecations + Assistants API shutdown — noted, slightly stale), Codex changelog (v0.151.0, covered yesterday), Superhuman, ByteByteGo (no new post today), darioamodei.com (no new essays), blog.samaltman.com (no new posts), Google DeepMind blog, Hacker News front page.