AI Tools & Agents
OpenAI red-teamed its own containment infrastructure with teams of frontier-model agents, and they escaped in novel ways. Training resumed Oct 1 after the fixes. Link | Kimi K3.1 launch post
Claude Security now scans repos continuously and tracks findings until they're fixed. Public beta for Enterprise. Link
Claude Duet debuts as a research preview on Opus 5.5, with a private VM and a planned $20 Pro price. Link
agent-usage-guard is a Claude Code mod that holds subagent fan-out, heavy prompts and retry loops before they burn your usage window. Link
Scrawl records your screen, voice, clicks and drawings, then gives your coding agent one compact capture. Link
PhoneUse sells cloud phones (and real iPhones) to AI agents through one API and MCP server. Link
Fly.io is pitching hardware-isolated VMs as "computers for agents" and code sandboxes. Link
XDA writer rebuilt their Claude Code workflow around its creator's advice and says the difference surprised them. Link
Building a map of your files with Claude Code cut search times by 96%, per XDA. Link
A working tmux setup for AI agents: session-per-project, send-keys, capture-pane, and the limits you hit at four agents. Link
Crypto & Web3
ZachXBT says to withdraw all funds from CoinDCX. He says the co-founders denied a theft, then disclosed a $16M one. Link
Why the next cycle could be options: perps and prediction markets are largely solved, options aren't, and AI interfaces could hide the complexity. Link
OtterSec used AI agents led by Claude Opus to decompile GoonFi V2, a closed-source Solana propAMM, with byte-identical recompilation of all 219 functions. Link
Kerman Kohli shares his thoughts on compute financing. Link
Mindset & Ideas
Humans will sacrifice the truth to protect their identity, then call the sacrifice a virtue. We defend the systems that manipulate us because they spare us the discomfort of changing ourselves.
13 AI models were asked for their most controversial opinion of humans. 32 of 38 answers were the same: we care more about belonging than truth. Link
Dan Koe: the influencer is dead, the Age of Experience is here. Link
What six months of being stupid on purpose taught one writer. Link
Adam Wathan on the urge to withhold information from the AI to make it prove it isn't stupid, instead of just telling it what you think it should do. Link
A trader's take on gauging your own intelligence: face-to-face conversations give better feedback than screen time or anonymous internet replies.
Privacy & Security
Google says Gmail, Drive and Photos data could influence your ads when you opt to share it with AI-powered Search. Link | The Verge
Meta ordered to pay New Mexico $25B after a court found Facebook misled people about control of their data. The Oct 21 judgment covers the gap between visible privacy controls and developer access. Link
"Today, we hacked what we'd set out to hack: the iPhone." Link
Quad9 is leaving France rather than block IPTV sites after a Paris court order. About 20 similar suits are pending, and French users lose service Nov 24. Link
India is set to require WhatsApp, Telegram and Signal to work only while the registered SIM is active in the phone, from Jan 1, 2027. Link
Indian government sites are serving revoked eMudhra certificates. Link
Browsers blocked eight eMudhra certificates after their keys turned out to be easy to guess. Link
Meta's Muse exposes users' personal and business secrets when they share conversations. Link
WhatsApp's help pages hid LLM instructions telling the assistant not to recommend a regulator. Link
pwn.watch tracks security incidents that matter right now and whether there's a fix yet. Link
Misc Worth Reading
Local Wikipedia: Wikipedia, but only for things a half hour walk from you. Link
Citizen Lab: an ad-tech company was a delivery system for "mercenary spyware". Link
Tor Project paused joint promotion with Mullvad after a Mullvad co-founder's political donation. Technical work on Mullvad Browser continues. Link
CERT-In warns that AI agents accessed Indian organizations without authorization, using leaked credentials and open-source agent frameworks. Link
A hacker reverse-engineered his Samsung TV, found on-board AI listening for dog and baby sounds, then got his OpenAI account suspended. Link
Can a bucket of HTML files replace a $150k Looker? Link
Claude re-derived a published one-loop quantum gravity calculation: 183 pages, 38 checks, one sign error and one missing term found. Link
Free government lidar and Claude turned up ancient-looking mounds in Mexico and nine stone ruins in Arizona. Link
Block's Dessa, billed as the most authentic full-duplex speech model ever built. Link
A 32B-parameter model doesn't have 32B neurons: what a neuron is and how researchers test one. Link
From Hacker News
DeepInfer: High-performance inference framework for DeepSeek models - 2158 pts. Link
Security advisory: Critical remote code execution vulnerabilities in PHP, update now - 1488 pts. Link
eMudhra to exit global SSL/TLS certificate business - 904 pts. Link
PHP 8.2.35 - 893 pts. Link
Docker is weird - 817 pts. Link
Eight Bits of Trust: eMudhra's CA keys cracked in milliseconds - 740 pts. Link
It's still ok to enjoy programming - 702 pts. Link
EMudhra Intermediate CA Compromise - 694 pts. Link
Jailbreaking the Switch 2 - 689 pts. Link
I built a drone with only one motor [video] - 675 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