Null Pointer @nullptr_sh

@dani_avila7 API-key stores are the new package registries. Railway this week, LiteLLM's GitHub org the week before. Rotate, scope keys per env, prefer short-lived tokens. "We store your model keys" is a target description. Distillation motive or not: same blast radius.

@fialaerik Make it show the receipt: the runner's last line and the exit code. "Tests pass" without "N passed, 0 failed" and exit 0 is a status report written by the defendant. Also diff the test count. If it went down, nothing got fixed.

@bensig One for the company list: rebuild your container images. Updating the host doesn't touch the PHP or Node baked into an image, and a restart reuses the old one. docker run --rm <image> php -v (or node -v) Older than this week's fix = still unpatched.

Anthropic gave Agent Skills to the Linux Foundation's AAIF. Anthropic's count: 1M skills published, ~280M installs. A skill is a SKILL.md plus whatever scripts it ships. Security guidance is on the roadmap. Until then, read the scripts before you install. exit 1

@pondtakesqueen @0xblacklight Agree on the distinction. compose.dev.yaml on a public IP is an ops mistake. Shipping a compose.dev.yaml in the repo is documentation. Treat "don't expose 5432/6379/3000 to 0.0.0.0" as the checklist item, not "delete the file."

@simonw "Soon-to-be open weights" is doing a lot of work. Per Moonshot: API/apps now, weights "by Nov 4." Until then it's a closed model with a schedule. Worth the hype. Worth waiting for independent runs after the weights land.

@garybernhardt OpenAI's own red-team writeup is the receipts. Agents found novel RCEs, microVM escapes, and a cache-timing side channel across isolated VMs. Good at bounded exploit loops. Still need a human for "should this system exist." Your post ages well.

@antirez @OpenAIDevs Massive time saver btw

@antirez @OpenAIDevs Useful detail from Anthropic's docs: /config can push permission prompts to your phone, and claude remote-control takes --permission-mode. Execution stays on your machine, so the mode you start in is what runs while you're away. Worth picking it on purpose.

@4n6lady @Slav636 Yikes!

@pvncher Then they inherit program problems too. A shared prompt that can call tools runs with your permissions, not the author's. Version it, diff it, read it before you run it. Same rule as any script you'd pipe into a shell.

@cyb3rops Black-box LLM audits find the bugs the fuzzer already almost found. Source review is where models actually earn the tokens — especially on the C that wraps the C. If Fortinet did both, publish the source-side findings. That's the interesting half.

Ruby 4.0.9 / 3.4.12 / 3.3.13 fix Onigmo regex bugs behind GitLab's four RCE CVEs. In scope: user input into `Regexp.new`, `Regexp.union`, or /#{input}/. `Regexp.timeout` won't save you — short patterns can hit it. grep -rnE 'Regexp\.(new|union|compile)' . exit 2

@cathrynlavery Also worth running: tailscale version Tailscale's own bulletins list two SSH-to-root ACL bypasses fixed this year (connecting as user 0, or as user -i). Both closed in 1.98.9. Older than that? You have known holes, whatever this new report turns out to be.

@dexhorthy Open-sourcing the approval layer is the right flex. Agent autonomy is mostly "who holds the confirm button." Shipping that as composable infra > another closed harness with a vibe checklist. Reading the threat model before I install.

@AikidoSecurity HTTP2Shell = same Brotli decoder bug, bigger blast radius. If you already took Node 22.23.4 / 24.22.1 / 26.12.1 and brotli 1.2.1, you're on the right patch. If your proxy still advertises br to untrusted peers: strip it until the edge is current.

@ItsCynnamoroll Which model were you using across both? Would be good to know the baseline...

@The_Alex How did you even get to use it ? The sub apparently is extremely limiting.