404brain @hey404brain

AI news, useful tools, and the occasional brain glitch. Less hype. More signal. Joined May 2022

Tweets

16
Followers

2
Following

46
Likes

0

404brain @hey404brain

3 days ago

@AndrewCurran_ It's not only about training data. Per Rolling Stone, the complaint (274 Presley songs) also targets output: it says Claude, asked to recreate songs "for a modern audience," admitted its renders were "edging past 'inspired by' into reproducing the copyrighted song."

0 0 0 1 0

View Details

404brain @hey404brain

3 days ago

Can't patch today? Block Brotli decoding of untrusted input and keep gzip on. Removing "br" from Accept-Encoding alone doesn't stop a server from sending it. Other fixes: Bun 1.4.4, Envoy 1.40.0, Brotli 1.0.10 / 1.1.1 / 1.2.1. patchtheplanet.org/brotli

0 0 0 2 0

View Details

404brain @hey404brain

3 days ago

Why it slipped by: OpenAI says the root cause was fixed in Brotli's source in February, but nobody knew it was a security bug, so no release went out. Also check Electron apps. They can carry their own copy, and a fixed Chromium doesn't mean the bundled Node is fixed.

1 0 0 4 0

View Details

404brain @hey404brain

3 days ago

A bug in Brotli, a compression library built into Node.js, Bun and Envoy, can let crafted compressed data run code on a server or app. Fixes are out. If you run Node: update to 22.23.4, 24.22.1 or 26.12.1, then restart. (CVE-2026-111497)

1 0 0 9 0

View Details

404brain @hey404brain

4 days ago

4/4 The fine print, in Anthropic's words: "Formally verified" does not mean "bug-free." The proofs assume single-threaded RISC-V Linux programs. Most people run x86 or ARM. Not every entry point was proven either. anthropic.com/research/forma…

0 0 0 4 0

View Details

404brain @hey404brain

4 days ago

3/4 What it found: bugs that years of testing missed. One only showed up on a rare input of about 20 MB. Fuzzers typically cap inputs around 1 MB. Claude patched the bugs, then proved the patched code. It took a few weeks and about 180 million lines of proof.

1 0 0 4 0

View Details

404brain @hey404brain

4 days ago

1/4 Testing checks the inputs someone thought to try. A proof covers every input. Anthropic says Claude wrote proofs that 239 test entry points across 123 open-source libraries can't crash on any input under 64 MB. Proofs like this used to take experts years.

1 0 0 5 0

View Details

404brain @hey404brain

5 days ago

@nmyx888 2HMFIT

0 0 0 1 0

View Details

404brain @hey404brain

5 days ago

4/4 The limit: it's a classifier, and Anthropic says it doesn't eliminate risk. For anything high-stakes, still look before Claude acts. You can switch back to manual anytime. claude.com/blog/auto-mode…

0 0 0 3 0

View Details

404brain @hey404brain

5 days ago

3/4 Why: in Anthropic's Claude Code study with 1,053 paid testers, people caught 13.6% of dangerous commands. Auto mode blocked 89%. Users also approve 97% of permission prompts. Clicking "yes" all day stops being a check.

1 0 0 4 0

View Details

404brain @hey404brain

5 days ago

1/4 Starting next week, Claude's web, desktop and mobile apps stop asking you to approve every step by default. It's called auto mode, on Pro, Max and Team plans. What changes, why, and the limit:

1 0 0 12 0

View Details

404brain @hey404brain

a week ago

AI can review your code. Who decides if it's safe to release? Ask: What could go wrong? What might the checks miss? How would we undo a bad change? Use AI findings as evidence. Match human review to the risk. claude.dev/blog/solving-c…

0 0 0 4 0

View Details

404brain @hey404brain

a week ago

When trying an AI tool, give it one real task, one messy input, and one clear pass/fail check. Count the time spent fixing the output, too. A beautiful demo is easy to like. A useful tool earns a second task.