AI Tools & Agents
Shopify let agents attack its own live infrastructure: tens of thousands of engagements, dozens of real findings, no outage. Link
ChatGPT Wallet lets agents pay without seeing your card, with a security check on every purchase. Rolling out to Pro subscribers in the US. Link
Anthropic upgrades Claude Code Projects to run parallel AI workers hands-free. Link
Fable 5.2 at High effort roughly matches Opus 5.5 at Max on FrontierCode 2.0 for about half the cost per task, and 1,000 Fable agents formalized Fermat's Last Theorem in 30 hours. Link
48 hours with Fable 5.2: burns limits fast ($500 in credits in 20 minutes on fast mode), but a monster at analysis and the lowest sycophancy he has seen. Link
Anthropic's Fable 5.2 prompting guide: tell the model "this codebase is yours" and it merges duplicate code far more often, at about 1.5x the output tokens. Review the diff, not the summary. Link
Claude Whiteboards, Claude Sheets and Claude Data are live on all paid plans. Link
Anthropic is seeing superexponential algorithmic progress, says a former employee citing a slide from the IPO roadshow. Link
OpenAI's Daybreak Red Team: coordinated AI agents that test web apps, validate reported vulnerabilities and verify patches. Link
Your agent can be turned. An approval prompt is written by the agent, so a hostile page it read an hour ago can help write it. It needs a handler, not a prompt.
Crypto & Web3
Brave Search API now accepts agentic payments via the x402 protocol. Link
Farcaster Articles are just chains of ordinary casts. Changing one changes what the network shows.
Mindset & Ideas
The Bitter Wall: the point where smarter AI stops producing better results, because the AI doesn't know what you want. Link
Patrick Collison asks what the positive AI scenario actually looks like: do people work, what is scarce, who is in control? Link
Have you considered Actually Trying? A LessWrong case for taking the effort to stop the AI race seriously. Link
Handing Over Humanity: even the optimistic post-AGI vision means ceding control of our future to a godlike AI. Link
The Antifragile Internet: agents make attacks cheap, but they can also harden the internet through automated security and agent identity. Link
The Jackpot Age: a coin that wins on average still takes almost everyone to zero if they bet too much, so size for the typical outcome. Link
Writing as a Thinking Tool. Link
The people not impressed by AI today have forgotten what AI video looked like 2 years ago. Link
Privacy & Security
One blanket Android permission just cost Meta โฌ805M. Italy fined it over Facebook's gallery suggestions, which reached photos of about 20M Android users without specific consent. Link
Claude Code's VS Code plugin shows a symlinked path as inside your workspace when it resolves outside it, so a prompt-injected agent could reach ~/.ssh. Reported unpatched as of v2.1.322. Link
GrapheneOS on the Pixel 11 is on indefinite hold because Google hasn't released the Android 17 QPR2 source to AOSP. Link
Ctrl+C, Ctrl+v, Ctrl+cve: bypassing Android's clipboard read security measures. Link
Electron, we have a problem: many apps auto-update without verifying signatures and still trust the compromised eMudhra certificates. Link
Terminal2Shell: researchers disclose a 55-year-old flaw affecting macOS, Windows and Linux, with no patch (per the disclosure site). Link
Signal fork Molly has had no updates since July 31, leaving users exposed to fixed Signal bugs. Link
From iCloud 27.3, "Anyone with the link" shares are readable by Apple, reviewed for "harmful content" and no longer end-to-end encrypted, even with Advanced Data Protection. Link
The Verge asked every TV maker 17 questions about how they spy on you. Link
A comprehensive study of how fingerprintable users are. Link
Football & Sport
United's only aerial threat is Harry Maguire.
Misc Worth Reading
AI Is Destroying the Trust Between Writers and Readers (The Atlantic). Link
A visual explanation of reinforcement learning. Link
How to create your own decentralized messenger protocol. Link
OSINT case studies and investigation examples from 12 cases. Link
Free Media Heck Yeah: a directory of free media. Link
Formalizing CPython: AI agents wrote 613,318 lines of Lean that copy Python's internals and turned up 418 bug reports along the way. Link
OWASP's Securing Agentic Applications V2: a builder and defender's guide covering architecture, identity, runtime protection and supply chain. Link
Open Source Says No: a catalogue of OSS projects banning AI-generated contributions. Link
A Disaster for U.S. AI Supremacy: the case that Jensen Huang's forced exit from Nvidia is a technological tragedy. Link
Open Kernel Challenge: the author asked the top 50 how they did it, covering agents, models, GPUs and spend. Link
From Hacker News
Jensen Huang resigns (3733 pts). Link
FAA issues national ground stop, lifted 4.5 hours later (1545 pts). Link
SpaceX catches the Starship to end Flight 15 (1307 pts). Link
Foldable Flight: The Best Paper Airplanes (1275 pts). Link
Nvidia is preparing to oust Jensen Huang (1220 pts). Link
Tim Peters has died (1205 pts). Link
Smart home devices should work without the internet (1160 pts). Link
Nvidia allegedly engineered scarcity by funding AI startups to buy its GPUs (1077 pts). Link
Service degradation on Microsoft 365 (1043 pts). Link
My AI agents found an unmapped Maya site? (990 pts). Link
That's a wrap. See you next Sunday. kazani
AIL 3: kazani picked every link; Claude wrote the one-liners and grouped them.
Daily reads on Telegram: Join here
Follow me on Farcaster: kazani




