A ten-person team at Anthropic claimed a three-year bounty by showing the NIST curve seeds came from drand48 seeded with the clock. Postgres slipped a month after AI code review found more defects than the humans could triage. And a Home Assistant update turned into a public fight about who opted you into what.

Today: the seeds, the slip, the scan, a 37-year-old OpenBSD bug, a $20 anti-cheat bypass, and a physicist’s satire that actually bites.

Defaults and deadlines tell you more than announcements.

🎲 The seeds were just timestamps

Three years ago Filippo Valsorda posted a bounty: explain where the seeds behind the NIST elliptic curves came from. The seeds shipped with the standard in 2000, provided by the NSA with no explanation, and the best lore was that the NSA’s Jerry Solinas had hashed a funny sentence and forgotten it. On October 15 at 01:07:51 UTC, a ten-person Anthropic team, including John Wood, Henry de Valence, Nicholas Carlini and Steve Weis, claimed the full bounty. They donated the prize to ISRG, the nonprofit behind Let’s Encrypt, which tripled it to $36,864.

The answer is almost disappointing: the seeds are output of drand48, the old POSIX random generator, seeded with the Unix timestamp at generation time in 1997 and 1998. Valsorda reimplemented the algorithm in Go to check the claim and lists which timestamps produced which curves. Anthropic’s write-up calls it a fun closing chapter and says nobody needs to change anything. Valsorda adds the security math: even if the NSA had picked any timestamp and any early candidate, it had about 30 bits of freedom, not enough to plant the weak-curve backdoor that Bernstein and Lange once sketched. All of this is the teams’ own accounting; we read both posts and ran nothing.

Valsorda’s post  ·  Anthropic’s write-up  ·  HN discussion

🐘 Postgres slips a month; the AI found too much

PostgreSQL 19 was expected in late September or early October. Instead the release target is about November 19, and eight to ten significant features were reverted to next year’s Postgres 20, including SQL Property Graph Queries, which Tom Lane had bet dinner would ship with unfixable bugs. Peter Eisentraut, a committer and VP at EDB whose own features were among those reverted, wrote up what happened on October 9 under the title "The reverts will continue until morale improves."

His account: some reverted features had real architectural defects. But the new pressure is LLM-assisted code review, which finds a long tail of relatively harmless issues fast. His image: you are staring at about forty defect reports, each a three-line fix, and the fixed overhead of each report eats the beta period. The features were written before LLM review was useful, so the quality bar rose after the code landed. He is explicit that this is not vibe coding, and says the cycle should sort itself out if the tools keep improving. A heavy security year he calls the Vulnpocalypse also pulled senior reviewers away. His claims, on his blog; we read it and The Register’s report and tested nothing.

Eisentraut’s post  ·  The Register report  ·  HN discussion

🏠 The scanner nobody opted into

On October 16 a user who writes as Viss, who says he runs security consultancy Phobos Group, filed an issue against Home Assistant 2026.10.1: he never authorized an integration that scans his secrets and sends anything up the line, and he wants it opt-in. Home Assistant founder frenck first replied that it is not a privacy invasion, pointing to the k-anonymity design, and closed the issue as won’t fix. Other users piled in, one calling the response dismissive. By late evening frenck apologized for being dismissive and posted the off switch: ha security options --pwned=false.

The feature is a pwned-passwords check against Have I Been Pwned, and frenck says it has been part of the Supervisor for years: hashes only, built for privacy, the same approach 1Password uses. Viss says he knows HIBP well and still objects to any outbound check he did not approve. Both can be true: a careful design, and a default nobody was asked about. We read the full thread; every position quoted here is the speaker’s own.

The issue thread  ·  HA on the check  ·  HN discussion

🐡 A bug old enough to run for office

Andrew Griffiths of security firm Calif reported a set of bugs to OpenBSD earlier this year. One traces back about 37 years to the original BSD sources: the shared-memory attach counter, shm_nattch, was a signed 16-bit short. Fork enough processes with shared memory attached and the counter wraps past 32,767 into negative territory, opening a use-after-free through a race. OpenBSD fixed it in errata 026 and 018.

The write-up walks the full path to local root with KARL, SMAP and SMEP all enabled: map and unmap 127 SysV segments, fork 259 processes to wrap the counter, groom the heap using malloc stats from sysctl, win the race, then use a -1 out-of-bounds write on a multicast socket option to flip a page-table entry from user to supervisor. No info leak needed, because OpenBSD’s kernel allocators are deterministic even while KARL randomizes the kernel image. He suggests mitigations: safe-unlinking, randomizing the page-table slot, and nondeterministic allocation. This is his research; we read it and ran nothing.

Calif’s write-up  ·  HN discussion

🎮 Beat the attestation with a $20 part

Epic announced version 2 of its Easy Anti-Cheat for Linux: eBPF monitoring anchored in secure boot, measured boot and TPM remote attestation, so the server can check your kernel is unmodified. Matthew Garrett, mjg59, a longtime Linux security developer, looked at the design and found the seam: the quote has to come from a real TPM, but not necessarily your TPM.

Two attacks, both his reasoning rather than a demo. On many desktop boards the TPM sits on a plugin module with an exposed reset pin; short it to ground and the TPM returns to its initial state, ready to accept whatever measurements userland writes. Or spend about $20 on a USB-to-SPI bridge and a spare TPM module, point the kernel at that one, and replay measurements from a clean boot; the modified kernel can simply lie to the eBPF agent. His verdict: remote attestation on PCs fails wherever you cannot bind a specific TPM to a specific machine. Epic could still ban TPMs server-side as cheat detection fires, which raises the cost. We did not build any of this.

mjg59’s post  ·  Lobsters discussion

📚 The Library of Babel, now with an IPO

In early October OpenAI published what Rikab Gambhir calls every mathematician’s childhood dream: a list of 722 write-ups of possibly-true mathematical facts. Gambhir, a theoretical particle physicist, answered with a guest essay on Terry Tao’s blog, and it is the sharpest thing written about the event. His opening reworks the famous confession: then they came for the physicists, and it was actually really cool, and they gave out free tokens.

His frame is Borges and the Library of Babel: a library holding every possible book already contains every proof, so by the research standards of 2026 it would be the most successful institution in history. The charge: generating probabilistically correct-looking text outsources the tedious, unprofitable work of checking and understanding to the community, and removes the old bottleneck that you had to understand a solution to count it. It is satire, clearly labeled, from a working physicist, on the blog of a mathematician who runs serious AI-assisted math projects himself. Tao notes the post was converted from another file format using AI, which is either irony or commitment to the bit.

The essay  ·  HN discussion

Mercor maps AI benchmarks onto the US economy. Its index of 29 benchmarks covers 208 of 923 O*NET occupations and 924 of 18,797 tasks; Computer & Math is best covered, healthcare barely. Mercor sells AI labor data, so the gap also argues for its product. Mercor Benchmark Map →

Follow-up: Muse Spark Open on Apple silicon. After No. 013’s lead, Apple’s ML team shows the model serving from Mac Studios with MLX distributed inference. Apple ML Research →

Follow-up: PS5LM doubles its stars. No. 003 covered llama.cpp on a jailbroken PS5 at 67 stars; back on HN this week at 380, GPL-3.0, still jailbreak-only. cobanov/PS5LM →

Follow-up: Gemini 4 Argon plays StarCraft. A Show HN from No. 011’s Argon model: live benchmark matches you can watch. StarSkirmish bench →