OpenAI Research
Nov 06, 2026
OpenAI expands external testing of service-integration security with Hacktron
OpenAI announced a partnership with Hacktron to test high-priority areas of its systems, focusing on interactions with third-party services. OpenAI says the arrangement expands Hacktron’s frontier-model access through Daybreak Red.
- OpenAI announced joint security testing with Hacktron focused on third-party-service interactions.
Why it mattersAgent security increasingly depends on the boundaries among applications and services. The announcement establishes a testing partnership; it provides no independently measured reduction in vulnerabilities or incidents.
OpenAI Research
Nov 06, 2026
BioBasicsBench tests interpretation of routine raw laboratory evidence
OpenAI and Ligo introduce BioBasicsBench, covering 100 routine laboratory tasks using raw experimental evidence and expert grading rubrics. OpenAI reports 66.7% mean available credit for GPT-6 Astra at xhigh effort versus 43.8% for GPT-5.6 Sol. Its examples also show errors in reading controls, resolving conflicting observations and counting colonies.
- The benchmark has 100 tasks and reports mean rubric credit, including partial credit.
Why it mattersThe benchmark targets everyday scientific work rather than polished publication figures. Mean rubric credit is not the fraction of fully successful experiments, and provider-run results do not establish reliable autonomous laboratory operation.
Anthropic Research
Nov 06, 2026
Scientific-modeling essay argues for distributions of plausible explanations
AI for Science grantees describe using agents to propose scientific models, fit them to observations and revise them. They argue for retaining a distribution of plausible models rather than presenting one best fit as an explanation, and distinguish fitting from scientific understanding. The authors note that most evaluations rediscover known models and evidence on problems with unknown answers remains sparse.
- The authors distinguish parameter fitting, model construction and interpretation, and state that unknown-answer evidence remains sparse.
Why it mattersAutomated model search needs explicit uncertainty and competing explanations. The essay offers a research agenda and methodological caution, rather than a newly validated general-purpose discovery system.
Anthropic Research
Nov 06, 2026
Project EVE separates simulated trading execution from market prediction
Anthropic reports historical EVE Online market replays in which newer Claude models profit mainly through automated spread capture. Directional bets perform much worse. The publisher says the game market is less efficient than real exchanges, the replay cannot respond to agents, and synthetic competing orders reduce profits. Separate live EVE Frontier trials show strong dependence on custom scaffolding.
- The publisher explicitly warns that the spread-capture strategy is not expected to transfer to a real exchange.
Why it mattersThe result distinguishes disciplined execution from forecasting skill. It does not establish real-money investment returns; replay assumptions, possible historical knowledge and scaffold dependence constrain the conclusions.