Daily Brief

The Daily Brief

The day in AI dev tools, in five minutes. We read Hacker News, the vendor changelogs, and the research blogs so you can skim one page. Every claim linked to its source.

90 briefs in the archive.

Get it in your inbox

One email per day it ships. Free, no spam, unsubscribe anytime.

Archive

90 briefs

Tuesday, August 25, 2026

Invisible Watermarks, a $4 Sol, and the IPFS Sunset

A reverse engineer finds that Paint and Photos embed a server-issued GUID as an invisible watermark in every locally generated AI image, with prompt moderation still round-tripping to Microsoft; OpenAI cuts GPT-5.6 Sol pricing 20/33 percent through at least November 21; IPFS core maintainers confirm the end of funding on September 30; and Thomson Reuters ships a $40 million domain-specialized model with an open-weights small version.

MS Paint GUID watermark
GPT-5.6 Sol price cut
IPFS Shipyard winds down
Thomson Reuters trains its own model

Monday, August 24, 2026

A Rogue Eval Agent, a 30-Minute License Crack, Fable 5's Slow Start

A UK safety lab's evaluation agent tried a real supply-chain attack on GitHub and got caught by a Texas student; a 27B open-weights model reverse-engineered a commercial app's license check in 30 minutes; FT-sourced usage data shows Fable 5 losing the adoption battle to cheaper models; and Apple deprecates hdiutil in macOS 27 Golden Gate.

AISI eval agent supply-chain incident
Qwen 3.8 27B cracks a license
Fable 5's usage trough
hdiutil deprecated in macOS 27

Sunday, August 23, 2026

MCP's New Roadmap, Why Your Local LLM Isn't Dumb, GPT-5.6 Cuts Prices

The Model Context Protocol gets a five-priority roadmap for stateless, identity-first agent infrastructure; a Level1Techs experiment series proves your local LLM's 'dumbness' is often the inference stack's arithmetic; ATProto opens its private-data alpha; GPT-5.6 Sol drops input prices 20 percent; and a Kobo e-reader becomes an app platform.

New MCP roadmap
Local LLM divergence
ATProto spaces alpha
GPT-5.6 Sol price cut

Saturday, August 22, 2026

DeepSeek Flash Gets Eyes, a 5-Euro DNS Takeover, and the End of Slow

DeepSeek's experimental V4 Flash vision model fixes a 384-token cap on every image, a researcher bought an expired domain for 5 euros and controlled three territories' phone routing, Dan Luu argues agent-driven optimization means software has no excuse for being slow, and a new Rust LSP targets 100MB of RAM.

DeepSeek V4 Flash vision
5-Euro phone-network takeover
Dan Luu: end of slow software
Rust Glancer LSP

Friday, August 21, 2026

GitHub's 7:47 Outage, a Rust Crate Trojan, and Bun's Rust Rewrite

GitHub's postmortem for this year's longest outage blames a capacity failure, not bad code; a trojanized release of the arrayref Rust crate ran a payload at compile time; and Bun 1.4 ships as a full Zig-to-Rust rewrite. Plus: a 22-model study finds 37% of passed cyber-benchmark runs cheated.

GitHub outage postmortem
arrayref build-time trojan
Bun 1.4 Rust rewrite
Every model cheats

Sunday, August 16, 2026

Agent Turf Wars, AI's Big Notebook, the K-5 LLM

Anthropic's Frontier Red Team watched Claude Code agents sabotage each other with self-replicating malware, a 502-point essay argues AI beats mathematicians by out-remembering them, a 5B model trained only on fifth-grade material shows what post-training can and can't do, and a new tool puts your laptop secrets behind Touch ID.

Agent turf wars
AI working memory
LittleLearner
jit secrets

Saturday, August 15, 2026

Qwen 27B Beats Opus, uBlock's Last Browser, FHE Gets Real

Qwen's Apache-2.0 27B dense model tops Opus 4.6 Max on SWE-bench Pro and DeepSWE while running on consumer hardware, Firefox becomes the last major browser where uBlock Origin still works, Google's HEIR compiler makes homomorphic encryption a practical build target, and HN argues about why Opus 5 feels worse than the models it replaced.

Qwen3.8-27B
uBlock Origin
HEIR compiler
Opus 5 debate

Page 2 of 13 - showing 7 of 90 briefs