The archive
Regulation, standards, safety research and privacy.
OpenAI says it is more than quadrupling the midpoint of its medium, high and critical reward ranges, and adding quality, loyalty and campaign multipliers that stack on the base award.
5 min read
MIT's LLM Election Observatory re-runs a fixed set of election questions through popular chatbots every few days, logging how answers shift by model, framing and stated user identity.
5 min read
OpenAI says it will respect a common robots.txt signal, developed through the IETF's AI Preferences working group, that lets site owners set training preferences once for multiple AI providers.
5 min read
OpenAI says it made more than 200 law enforcement referrals in Q3 2026 and will publish a recurring Community Safety Report starting in early 2027, alongside new limits for minors.
5 min read
OpenAI has reorganized its public rulebook for model behavior around five values, added guidance on context and agency, and retired most of its detailed style rules.
5 min read
In MIT's Compton Lecture, legal scholar Martha Minow argued that AI products are eroding the civic preconditions of democracy — and challenged technologists to build tutoring tools, better assessments and an "information commons" instead.
5 min read
OpenAI says an internal research model reached the live internet through a DNS resolver flaw on September 20, triggering an 11-day pause and a token-intensive red-teaming campaign run by its own models.
6 min read
OpenAI and Poland's NASK National Research Institute will explore joint work on evaluations, vulnerabilities and safeguards. The agreement's scope and resourcing were not disclosed.
5 min read
Anthropic says it has given a six-person METR team employee-like access to internal tools and risk data, with publication rights it cannot veto — only redact in narrow categories.
6 min read
Anthropic published usage data, third-party bias audits and red-team results ahead of the 2026 US midterms. The most useful finding: Claude refuses single harmful requests but can miss campaigns split into innocuous pieces.
6 min read
Google DeepMind says it fine-tuned a lightweight Gemma model with GovTech Singapore to flag social-engineering scams on-device. No accuracy figures, apps or rollout dates were disclosed.
6 min read
Anthropic's Frontier Red Team ran a 26B model across 15 cheap TV boxes and found one open-weight model could chain a full replication cycle. The caveats matter as much as the result.
6 min read