Where access stopsAI Safety
OpenAI separates research access from live internet in frontier training
An offline collection of web material lets models research inside a controlled environment while external connections remain outside the boundary.
Find your next discovery
AI Safety
Agent security, cybersecurity, privacy and AI policy.
Where access stops
Training Data TestsAI Safety
Anthropic tests the gap between training-data screening and model behaviour
An automated search found data that could pass a monitor yet change a research model’s behaviour. The experiment examines two different layers of evaluation.
Connected data and adsAI Safety
Google’s connected services: Follow the data permission to its destination
The privacy policy links cross-service personalisation to the settings of the service receiving access. Where data is used matters alongside where it is stored.
A hot-wallet breachAI Safety
Financial access for AI agents: Reading information is a separate permission
A signing-infrastructure incident raises a useful automation question: how should observation, recommendations and real transactions be separated?
Discord on the WebAI Safety
Discord visibility for AI communities: Separate messages from summaries
Discoverable servers are gaining web pages. Community managers need distinct decisions for individual messages and broader channel summaries.
Private AI workflowsAI Safety
OpenAI Private Intelligence separates retention, processing and agent permissions
Enterprise privacy involves what remains after a request, how information is processed and what an agent can do with access.
Open scienceAI Safety
AI-assisted physics puts research access on the agenda
OpenAI describes different review routes across mathematics, computer science and physics. An independent working group is examining the physics materials first.
A new cyber thresholdAI Safety
GLM-5.3 research separates cyber capability from safeguard behaviour
Anthropic examines technical outcomes and responses to harmful requests as distinct questions, using isolated evaluation environments.
Coordinated PHP fixesAI Safety
PHP response shows the work between an AI finding and a deployed fix
PHP maintainers led remediation after findings from Calif and Wiz. OpenAI describes the coordination needed before and after disclosure.
Better care?AI Safety
Warhound combines genetics and digital equipment in a working-dog programme
The company announces $200 million in seed funding, with Colossal biotechnology and Anduril systems supporting a planned US breeding and training programme.
Who checks?AI Safety
METR examines OpenAI’s internal agent-monitoring chain
The July–August exercise studies historical coverage and the treatment of conversation-compaction events.
Risk or fact?AI Safety
Anthropic reports unintended model actions and changes to internal testing
The October 9 account groups four behaviours and describes suspending live internet access in internal evaluations while controls are verified.