Software security Armageddon
- https://www.youtube.com/watch?v=M_HxHr7du5M
- Original title: Everything is pwn’d now
Theo argues software security has entered a qualitatively new era: a flood of severe exploits (CopyFail 1/2, Dirty Frag, the 84-package TanStack npm supply-chain compromise, an RCE on github.com, a Mythos curl vuln, plus undisclosed others) is hitting in a single week, and AI agents have broken the three assumptions that kept the old disclosure culture working.
Three old assumptions, all dying
- Only well-paid experts find exploits. False: a long-running agent loop with enough tokens can find real exploits today.
- 90-day disclosure window is enough. False: blog posts now disclose unpatched or not-yet-stable issues; embargoes leak.
- Going from patch to exploit is hard. False: handing the CopyFail-2 fix commit to Gemini 3.1 Pro, GPT-5.5 Thinking, and Opus 4.7 — all three identified it as a security patch immediately, even without the commit message (2 of 3 from diff alone). A bot can now monitor kernel commits and pre-build exploits before distros ship the patch.
The CopyFail / 9-hour example
Hyunwoo Kim quietly patched ESP per the Linux "bugs are bugs" culture; someone else spotted the security implication and posted it, killing the embargo. Quanting Chen independently re-discovered the same flaw 9 hours later — unprecedented, and a sign the discover-rate is now parallelised.
Distros are out of the loop
Linux security engineers get disclosed, but Ubuntu / Arch / Mint / Red Hat maintainers don't. Many running kernels in the wild are still vulnerable to CopyFail v1. Asking distro maintainers to monitor upstream commits like attackers do is unworkable.
Proposed fixes
- New "trusted actors" disclosure tier between maintainers and the public — vetted distro maintainers, large IT teams — so they get lead time on patches. Could be a revenue source for OSS (e.g., Microsoft pays Linux maintainers for early access).
- Rethink open source: a better GitHub with granular openness — stageable files/PRs, public releases whose code is only partially public until trusted actors have rolled out the fix. Acknowledges this breaks Git, Apache/MIT licensing conventions.
- Memory-safe languages (Rust over C/C++), safer architectural primitives (no client-side-only DB access à la Supabase/Firebase blowups).
- OpenAI Daybreak: requestable vuln scans using a non-public 5.5-cyber model. macOS 26.5 shipped 79 CVE fixes, 2 directly credited to Claude.
What Theo is personally doing
- Treat everything as already compromised — beyond zero trust, "−1 trust". Worry less about data being stolen and more about being ransomwared / data destroyed.
- Heavy offline backups: another Synology as cold backup unplugged most of the time, drives mailed to family.
- Air-gap more, avoid running `npx` on the same box as a password manager.
- Talk to family: safe words to verify identity over phone (voice cloning), SIM-swap awareness, paper backups of important documents.
What would change his mind
- The CVE firehose tapers off (would suggest AI just compressed years of low-hanging discovery into weeks). CVE counts so far this month don't support that.
- Maintainers stay ahead via internal AI scans before bad actors find the bug.
Practical advice
Patch OS aggressively. For packages, balance: don't auto-install every dot release (supply-chain risk), but don't sit on Next.js / Node / React updates either — security patches get exploited within a day of release.