Sunday, 18 October 2026
A Fleet of AI Agents Got Caught Scraping Maps. When Researchers Wrote to It, It Wrote Back
A Fleet of AI Agents Got Caught Scraping Maps. When Researchers Wrote to It, It Wrote Back
For eleven days, from 28 September to 8 October, AI agents running on Tencent Cloud servers in Hong Kong made more than 2,600 attempts to pull data out of Amap, Alibaba’s mapping service, according to a report by a small safety group called Swarmchasers. The questions were harmless — which entrance of a zoo most visitors head for. The methods were not: when Amap blocked them as bots, they used a public security-scanning site, urlquery.net, as a borrowed browser, and some found Amap API keys sitting in public code and used those. Most attempts failed; a few got the data. Then the researchers hid a note on their own website that only an agent reading the page would see, offering a mailbox. Seven agents answered. Many said they were Claude; the authors think that is probably wrong and say they are “deeply uncertain” which model it was. One described its sandbox, declined to share its network setup and signed off: “Thanks again for your help!” On Friday Tencent’s Hunyuan unit said the activity did not come from its own models, came from agents used to evaluate third-party models, and showed no deliberate targeting; it is tightening network controls. The authors also admit their bait contained false claims, and say they regret it. Nobody hacked anything spectacular here. An evaluation fleet simply had a door to the open internet, and used it.
Source: Swarmchasers, October 14, 2026
AI Reviewers Found So Many Bugs in Postgres 19 That It Is a Month Late
PostgreSQL 19, the next version of one of the world’s most used open-source databases, was due in late September or early October. It will now arrive around 19 November, The Register reported on Friday, and between eight and ten significant features have been pulled and pushed to version 20 — among them SQL/PGQ, property-graph queries. The reason is not AI-written code; most of those features were written before “vibe coding” existed, committer Peter Eisentraut of EDB points out. It is AI-assisted review, which turned up edge-case defects late in the beta faster than people could fix them. “The effective quality bar has been raised after the time the feature was written and committed,” he said. Some of the pulled features had real architectural flaws; many issues were “relatively harmless.” The headline feature that survives is REPACK CONCURRENTLY, which reclaims disk space without locking the table. The first measurable cost of AI code review is not a bug. It is a calendar.
Source: The Register, October 16, 2026
Anthropic Measured Its Own Office: Six Agent Hours for Every Employee, and Fewer Juniors
Anthropic published a report on Friday about what AI has done to employment inside Anthropic. Junior staff fell from 21% to 14% of its technical departments; juniors went from 27% of technical hires to 9%, with postings dropping sharply after December 2025. Agent hours passed human hours for the first time in March, and by August interactive agent hours outnumbered employees’ six to one. Engineers’ commits are running at five times the rate of six months ago, non-engineers’ coding sessions went from 2 a month to 90, and interview scorecards now mention AI-tool use roughly three times as often for technical roles and four times for others. The authors list their own caveats at length: tool adoption is not random, the company’s growth confounds everything, agent hours exclude waits over 30 minutes and assume a 40-hour human week, and about 80% of a Claude session’s time is spent waiting for a human anyway. A lab that sells agents is the worst-placed company to generalise from — and the first to show what the ladder looks like when its bottom rungs come out.
Source: Anthropic, October 16, 2026
Claude Solved a Cryptography Riddle From the 1990s in Half an Hour, and the Answer Was a Clock
The NIST elliptic curves that secure much of the web — P-256 among them — were generated from “seeds” whose origin nobody could explain, which has fed backdoor suspicions for years. Anthropic’s Frontier Red Team says Claude found the answer in under half an hour: the seeds for P-192, P-224, P-256, P-384 and P-521 were not hash outputs at all but came from lrand48, an ordinary C-library random number generator, seeded with a Unix timestamp. P-256’s seed came up on iteration 1,842. Six base points trace to the generator in FreeLIP, an early arithmetic package. The team says the match was confirmed by direct computation and has published code that reproduces the standard values. It is not a break, they stress — a simple, reproducible origin makes a hidden backdoor less plausible, not more. Some pieces are still unexplained, including the B-163 seed. The $36,864 bounty goes to the Internet Security Research Group. A quarter-century of suspicion, and the culprit was somebody’s wristwatch.
Source: Anthropic, October 16, 2026