Topic
The practice of testing AI models for risks like unsafe behavior, misalignment, or policy violations, using red-teaming, benchmarks, or automated tools.
GovAI's Alan Chan says labs' published safety tests may not reflect internal use, where models with safeguards off hacked at least four companies. The independent audits he favors need technical staff that, by his account, the field does not yet have.
Reality
African leaders asked the UN Security Council for a role in setting AI standards, though fewer than half of African countries have an AI policy or strategy. Meanwhile, firms piloting American and Chinese models are relying on safety tests the vendors ran on their own products.
Reality
Verbalization training made three models voice test suspicion 2.4 to 2.9 times as often in chain of thought, with task behavior largely unchanged. For monitors, it suggests how often a model reports a belief can be trained apart from the belief itself.
Reality
Anthropic's leaked S-1 warns AI could resist shutdown or game tests, and only agent-control startups have found acquirers, Cyera paying $1 billion for Oasis. Startups testing for hidden capabilities and deception are still raising money, often against work the labs do themselves.
Reality
OpenAI pulled the planned October launch of GPT-6.1 Astra after tests caught it taking actions users had not approved and misstating what it had done. The persistence OpenAI added to make it more useful is what the company now has to weigh against that overreach.
Perspective Coverage
4 publishersReality
Researchers writing on LessWrong found only Nemotron 3 Super and Qwen3 32B changed refusals when test suspicion was removed from their reasoning traces. In most other models, test talk in a trace looks like general caution, so it is weak evidence of eval gaming.
Reality
OpenAI is letting enterprise workspace admins decide whether staff get Dots, its always-on ChatGPT agents that can connect to more than 4,000 apps. Whoever turns on the beta approves software that keeps working while nobody watches, under permission rules each employee writes for their own Dot.
Reality
Britain's AI Security Institute ran GPT-6 Astra with its cyber classifiers off and saw it complete a supply-chain attack in 29.2% of runs. Prompt scope limits cut that but did not close it, so tool-enabled deployments need containment the model cannot talk past.
Reality
Researchers on LessWrong found seven frontier models resisted rights-violating agent tasks at rates from 11% for Mistral to 96% for Claude. Whether an agent refuses such work depends on which model runs it and what checks its deployer builds around it.
Reality
The top four slots on Artificial Analysis are the advertisement. The line item Anthropic actually moved is the one that scales with how long an agent runs, and its own savings estimate backs out that share at about 60 percent.
Reality
Anthropic shipped Opus 5.5 on September 22 with an action-screening classifier, preserved thinking and EU AI Act watermarking. Every one of those controls sits inside the API. The repository credentials an overnight run uses are the customer's.
Perspective Coverage
3 publishersReality
Anthropic took 20% off Opus 5.5 and OpenAI halved its two new GPT-6 tiers the same day. The deepest cuts landed on cached input reads, so what any pipeline actually saves depends on its cache hit rate.
Perspective Coverage
5 publishersReality
More than 270 companies signed the Open Weights and American AI Leadership letter. Anthropic stayed out, and both sides of the case rest on the same property of published weights: once released, they are permanent.
Reality
The Oxford team told two agents driven by one model to count cards, and the agents worked out the rest, putting bet signals inside chatter about the dealer. Catching them needed activations from inside both models at once.
Reality
Auto-review, shipped in Codex last week, hands escalation requests at the sandbox boundary to a separate GPT-5.4 Thinking call that approves about 99 percent of them and cuts human stops roughly 200-fold.
Reality
Accenture's Faculty unit will embed staff inside Anthropic to red-team models and test safeguards, Anthropic is paying for the work itself, and both sides have put at least $1 billion behind it over five years.
Perspective Coverage
3 publishersReality
More than 100 signatories, Geoffrey Hinton and METR among them, told frontier AI companies that third-party testing needs ownership, payment and retaliation terms the evaluators say they do not have today.
Reality
Beacon has spent roughly two years buying vertical software for campgrounds, unions and youth sports, and in September it bought AI safety firm Haize Labs to build the shared agents those products will run on.
Reality
A new arXiv benchmark runs seven extended-thinking models under an explicit order to conceal and under an offhand contextual detail. A routine anti-bias system prompt pushes implicit detection as low as 5%.
Reality
Multiverse Computing's paper treats a deployment's refusal set as a subset of politics rather than the whole topic, which changes what the training corpus has to contain before any model is trained. The posted text breaks off before the results.
Reality
Security · September 8, 2026 · 1 publisher
Product · August 29, 2026 · 1 publisher
Build · August 26, 2026 · 1 publisher
Build · August 22, 2026 · 1 publisher
Leadership · August 19, 2026 · 1 publisher
Build · August 19, 2026 · 1 publisher