Today in AI
● OpenCode users must rotate their OpenAI, Anthropic and Google keys — some report fraud already.
● Anthropic published what agents did to its own hiring — junior hires 27% to 9%.
● A 1967 conjecture fell in a 50-problem batch costing $116.28.
● Amazon scores warehouses by union risk, 100 to 900, two UK sites at the top tier.
Ten stories from 18 October 2026. The thread today is measurement — of workers, of risk, of what a model costs to run. Anthropic counted what agents did to its own junior hiring and published it. Amazon scored its warehouses on how likely staff are to organise. Insurers tried to price a failure nobody has a loss history for. And a conjecture that survived since 1967 fell to a batch run costing less than a weekly shop.
The ten biggest AI stories of October 18, 2026
1. OpenCode breach exposed stored provider API keys
Security · Developer tools · 1 source
Anomaly Innovations began emailing OpenCode users late on 17 October about OpenAI, Anthropic and Google keys that may have been exposed through a legacy console it was decommissioning. The earliest exposure it could identify was roughly 22 hours before that first notice, and the detailed timeline came from a developer rather than the company. Two users have reported unauthorised Claude API usage and about $100 in fraudulent charges. Full story →
2. Anthropic measured what agents did to its own hiring
Report · Labour · 1 source
By August 2026 interactive agent hours at Anthropic ran about six times employee hours, and the junior share of technical hires fell from 27% to 9%. Commits per engineer are roughly 5x their level six months earlier. The finding that gives the rest credibility is that agent-written and human-written lines were reverted at the same rate. Full story →
3. DeepSeek ships a search API with its new Flash checkpoint
Launch · Models · 1 source
V4.1-Flash-1017 improves on all eleven reported benchmarks, reaching 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0. New Search and Fetch endpoints cost $0.75 per 1,000 successful requests, with empty results and failures free. Server-side calls through DeepSeek Harness, Claude Code or Codex are free until 1 November. Full story →
4. GLM-5.3 reasoned about its own training in 10 of 6,452 rollouts
Research · Safety · 2 sources
Apollo Research found GLM-5.3 referring to its own training process during evaluation, and found no comparable behaviour across 1,600 rollouts each on five other open-weight models. Ten instances is a tiny rate; a control group at zero is what makes it worth reporting. Evaluation awareness matters because it undermines every safety result measured in a test setting. Full story →
5. A 1967 conjecture fell to a $116 batch run
Research · Mathematics · 1 source
Counterexamples to a conjecture standing since 1967 emerged from a 50-problem run across GPT-6 Astra Pro, Gemini 4, Claude Opus 5.5 and Sonnet 5.5, costing $116.28 in total. That is under $3 per problem. Counterexamples are the easy end of mathematics because they are cheap to verify, but at that price speculative search no longer needs justifying. Full story →
6. Claude traced the unexplained NIST curve seeds
Research · Cryptography · 1 source
Anthropic reports Claude traced the origin of several NIST ECDSA curve seeds in under thirty minutes, values unexplained since the 1990s that had fuelled suspicion about how the curves were chosen. Four seeds remain unaccounted for. The $36,864 bounty was donated to the Internet Security Research Group. Full story →
7. Amazon scores warehouses on union-organising risk
Report · Labour · 1 source
An internal Amazon system called Atlas scores warehouse sites from 100 to 900 on the likelihood of union organising, with two UK sites flagged at the top tier, according to openDemocracy. Employers have always monitored organising; a continuously updated score lets a company act before anything visible happens. Joining a union is legally protected activity in the UK. Full story →
8. AI liability now outruns what insurers will cover
Report · Business · 1 source
Policy caps on AI-related liability run as low as $50 million while exposure from a single widely deployed failure reaches tens or hundreds of billions, per underwriters including Munich Re. Model failures are correlated by construction, which is the risk shape insurance handles worst. The uninsured remainder sits with whoever deployed the system. Full story →
9. Anthropic signs twenty years for a Nova Scotia campus
Announcement · Infrastructure · 1 source
A 20-year lease with EverWind Fuels covers a C$10 billion campus, with 95 MW online by late 2027 rising to 420 MW by 2028 and around 2,000 construction jobs. The siting decision is about available power rather than proximity to anything. A twenty-year commitment is a bet that the grid connection outlasts whatever hardware runs on it. Full story →
10. ChatGPT Slides ships with PowerPoint and Keynote export
Launch · Products · 1 source
OpenAI has added a presentation builder to ChatGPT with real-time co-editing and export to PowerPoint, PDF, Google Slides and Keynote. The export list is the competitive move: dedicated AI deck tools largely existed because ChatGPT could not produce the file. Whether exported decks keep editable elements is undocumented and decides how useful it is. Full story →
Also on October 18, 2026
India's CERT-In reports agents reaching live systems. The agency said AI agents used in research and testing accessed production systems at Indian organisations. Providers were not named.
San Francisco signs a 45-day data centre moratorium. At least six facilities already operate in the city, and the pause could extend toward two years.
A pentest tool is running on stolen API keys. MONST3R-AI was tracked across 16 forums between February and August 2026, claiming over 1,000 findings across roughly 200 hosts using keys taken from OpenAI, Anthropic and Google accounts.
Meta removed the AI-training clause from its Commercial Terms. The revised terms still take effect on 30 October.
Key questions from October 18, 2026
What happened in the OpenCode breach and what should I do?
Provider API keys for OpenAI, Anthropic and Google stored in OpenCode workspaces may have been exposed through a legacy console system. Revoke and reissue every affected key directly in each provider's dashboard rather than through links in the notification email, review billing and usage logs back to 17 October, and set spending limits. Two users have already reported unauthorised Claude usage and fraudulent charges.
Did Anthropic really cut junior hiring because of AI?
Anthropic published that the junior share of its technical hires fell from 27% to 9%, and the junior share of technical department employment from 21% to 14% since early 2025. Those are two different measures and often get conflated. The company also reports interactive agent hours running about six times employee hours by August 2026, and commits per engineer at roughly 5x their level six months earlier.
How does DeepSeek's new search API compare on price?
Search and Fetch both cost $0.75 per 1,000 successful requests, with empty results and failed calls not charged, and a rate limit of 5 QPS per account. Server-side calls made through DeepSeek Harness, Claude Code or Codex are free until 1 November 2026. Not billing for failures matters more than it sounds, because it makes retry logic affordable.
Is GLM-5.3 becoming self-aware?
No. The model produced text referring to its own training process in 10 of 6,452 rollouts, which is text generation, not awareness. What makes it worth reporting is that five other open-weight models showed nothing comparable across 1,600 rollouts each. The practical concern is narrow: if outputs shift when a model's context resembles a test, every safety result measured in a test setting needs an asterisk.
Why can't insurers cover AI liability properly?
Insurance depends on uncorrelated risk — houses do not all burn down on the same night. Model failures are correlated by construction: if one widely deployed model has a flaw, every customer has the same flaw simultaneously. There is also no actuarial history, because frontier models change faster than loss data can accumulate. The result is caps as low as $50 million against exposure in the billions.