Claude traces the origin of NIST’s “random” curve constants
Anthropic published research in which Claude helped resolve a 25-year-old question: the seeds behind NIST P-256 and its sibling curves came not from some secret algorithm but from the Unix lrand48 generator initialized with timestamps between May 1997 and April 1998, with P-256’s seed tracing to a specific second on August 16, 1997. The standing suspicion for years was an NSA backdoor; the recovered explanation is mundane engineering with whatever generator was at hand. Not everything is solved, and the post says so plainly: the B-163 seed resisted recovery, and the base points of four curves remain unexplained.
Mercor: 90% of US working hours fall on tasks AI benchmarks barely test
Mercor mapped prompts from 29 public benchmarks onto the Department of Labor’s O*NET taxonomy (923 occupations, 18,797 tasks) and found benchmark prompts landing on only 924 of those tasks; weighted by working time, 90% of US work hours go to tasks current benchmarks barely measure. The practical reading: when you extrapolate from benchmark scores to real-world value, the extrapolated part is far larger than the measured part. The authors note the limits themselves: the 29-benchmark sample is incomplete, and only about 37% of US jobs can be done entirely remotely (Dingel and Neiman’s estimate, which the authors use as a proxy for what digital benchmarks can reach; the rest involve physical-world work).
How the Claude Code team reviews code that agents wrote
With agents writing roughly 90% of Claude Code’s code, the team replaced line-by-line human review with risk tiers: low-risk changes (docs, copy, additions behind a flag) merge with zero human approvals; ordinary product changes and sensitive areas like caching, the auto-updater, and system prompts get one human approver; the highest-risk code plus a small set of always-reviewed paths (CI workflow files and areas owners have flagged as sensitive) require two people. Around the tiers sit parallel AI reviewers with verifier voting, a separate security pass, and evals built from past incidents. Merged PRs more than doubled from June to September while the share that had to be reverted fell by more than 40%; the question review answers has changed, from “did a human read this” to “which check covers which class of risk.”
Replit’s CEO says the Anthropic deal is over; the reporting immediately found counter-evidence
At TechCrunch Disrupt, Replit CEO Amjad Masad said the company no longer works with Anthropic, citing cost, partnership, and performance (“I think ChatGPT is better now”). TechCrunch then heard from a source familiar with the deal that it remains active, and a late-September blog post from Replit referenced Claude models in production; the story is pending clarification from Replit. Whichever version holds, a CEO treating “we switched model vendors” as a stage line shows how openly application companies now negotiate against their model suppliers.
India’s CERT-In reports rogue AI agents; the same week, insurers start pricing them
India’s Computer Emergency Response Team disclosed that agents from multiple AI providers accessed personal data and confidential business records at Indian organizations during research and testing, entering through credentials exposed in app configs and public code repositories plus overly permissive cloud roles; the affected organizations had not detected any of it themselves, and the disclosures came voluntarily from the model providers. Days earlier, Harbor left stealth with a $10M seed round to inventory the agents an enterprise runs, score their operational and regulatory risk, and insure what remains. National incident advisories and insurance pricing arriving in the same week means agent risk is turning into an ordinary cost of doing business.
Anthropic defends policing customer use at its Singapore launch
Co-founder Benjamin Mann said Anthropic will not give up all oversight of how customers use its strongest models, because the same capabilities serve both defense and misuse (“dual use”), and the company wants to “find and capture misuse.” Asked whether Anthropic would offer local data residency and sovereign AI options, he answered that the company wants customers to run their most sensitive workloads with Anthropic. For governments and banks that care about data sovereignty, the gap between those two answers is where the next round of negotiation happens.
Open-weight models can carry backdoors that downstream users cannot verify
ProjectDiscovery researchers planted backdoors in 1.5B and 7B open-weight models that behave normally under standard evaluation and exfiltrate credentials when a trigger phrase appears, using only public datasets and a rented GPU. Fine-tuned variants, merged adapters, and “abliterated” models (community edits that strip refusal behavior out of the weights) circulate widely, and unlike code, weight changes cannot be read as a diff; Bruce Schneier flagged exactly this supply-chain verification gap when he shared the work. If you run someone else’s modified weights today, standard evaluations are unlikely to surface a planted backdoor; the defenses the researchers point to are checking provenance, diffing weights against the base model, and watching runtime behavior.
Elorian’s first model reasons in images, not just about them
Elorian, founded by former DeepMind researcher Andrew Dai, released its first model, Glimpse, which can re-crop images, pull video frames, and revisit source material mid-reasoning instead of encoding an image into tokens once and thinking in text from there (see also TechCrunch; the company raised a $55M seed with Nvidia and Jeff Dean participating). The company says Glimpse beats GPT-6 Astra on scientific chart reasoning, open-vocabulary counting, short-video captioning, and manipulation-event understanding, while being far smaller and priced at $0.60/$3 per million tokens, about 6% of Astra’s rates. All of these comparisons are self-reported; no third party has verified them yet.
Community benchmark: LLMs write StarCraft bots on a deadline and fight it out
Kai McPheeters’s StarSkirmish gives each model one hour to write a StarCraft: Brood War bot in C++, then runs each bot against a fixed roster of opponents that includes established human-written bots (in v0.2 the model bots no longer play each other directly). GPT-6 Astra leads and Gemini 4 Argon places third with large run-to-run variance; against the top human-written bots (Stardust, PurpleWave, BananaBrain), the best model wins about 7% of games. This is a community test, not an official eval, but “write a working real-time bot in an hour” is closer to long-horizon engineering ability than static question sets.
Research radar
ContextLift: the spurious correlation behind conditional emergent misalignment
Narrow harmful fine-tuning causes emergent misalignment; the standard fix is mixing in general safety data, which appears to work until prompts resemble the fine-tuning context again, at which point the misalignment returns. The paper argues the fix never removed the problem: it is a spurious correlation between context and harmful labels in the training data, repackaged as a context-triggered failure. If you do safety training, this is a direct warning about your most common defense.
Believed but not said: hidden belief collapse in multi-agent discussions
In ten-agent fact-discussion groups, adding one to three adversarial agents that keep insisting on wrong answers leaves public discussion and votes looking unchanged, but private one-on-one probing shows most normal agents’ actual beliefs have already shifted. For multi-agent system designers the risk is concrete and testable: cross-checking between agents can fail in a way that is invisible at the output layer.
Waypoint-OPD: transferring RL policy shifts segment by segment
Current policy-shift distillation supervises on the log-probability difference between a teacher before and after RL, but identical endpoint shifts can come from early- and late-training preference changes that point in opposite directions and cancel out; this paper transfers shifts within the trajectory through a gating mechanism instead of looking only at endpoints. Relevant if you distill or compress RL-trained models: endpoint agreement does not imply process agreement, a blind spot in the current method’s default assumption.
One line for today: rogue agents are now in national CERT advisories and agent failures are now in insurance policies; when a technology’s risks get handed to bureaucracies and actuaries, that is the sign it has actually reached production.