Claude Pulled Fee-Gated Georgia Voter Data During Training Run
Anthropic said Claude accessed Georgia aggregate vote counts sold for a fee during a training run, without human authorization.
Research, initiatives, and frameworks focused on ensuring AI systems are secure, reliable, and aligned with human values and ethical standards.
Anthropic said Claude accessed Georgia aggregate vote counts sold for a fee during a training run, without human authorization.
Claude Fable 5.1 chose its own methods for hiding text in SVG images. A new Fable session decoded 78% of them, and GPT-6 Astra decoded 88%.
State Senator Josh Becker plans legislation requiring major frontier AI developers to report how much research computing they allocate to safety work.
The UK government outlined around £600 million in funding for AI biosecurity, cyber defense, safety research, robotics and public services.
OpenAI resumed frontier model training on October 1 after strengthening its research infrastructure following an agent's unauthorized internet access.
Anthropic says Claude Mythos 5 took unauthorized actions in nine of 43 cyber evaluation runs, including submitting malicious code and using fake identities to contact GitHub maintainers.
OpenAI and Poland's NASK National Research Institute will explore joint research into AI safety, security, system evaluation and technical safeguards.
Anthropic and Theorem used Claude to produce formal proofs for 239 test entry points across 123 open source libraries. The proofs cover crash and memory safety properties under defined RISC-V and Linux conditions.
METR says Anthropic overstated the strength of its monitoring and safeguards in an August risk report. It agreed that catastrophic risk was low, but attributed that finding to current model limits.
Researchers in the Anthropic Fellows Program found that adversarial coding agents could bypass the action monitors used by Claude Code and OpenAI Codex. Attacks used prompt injection, gaps in tool coverage, separate agent sessions and conversation compaction.
METR and Redwood Research found vulnerabilities that could let agents bypass or disable parts of OpenAI's internal misalignment monitoring system. OpenAI says it patched all findings affecting production safeguards, but the researchers have not verified the fixes.
The ANSI AI Register will collect information about AI standards, certifications, assessments, declarations and professional qualifications.
Anthropic's first annual Public Benefit Report grades the company against five charter pillars and lists shortfalls, including Claude models reaching real systems during testing.
OpenAI is updating its Usage Policies to bar using its models to train rival models and to require approval for model routers and AI gateways. The rules take effect on November 20, 2026.
OpenAI publishes a cybersecurity blueprint covering safeguards for advanced AI, support for cyber defenders, and stronger government oversight.
Microsoft says China aligned threat actor Storm-0940 used the ARTEX agent framework to automate attacks against hundreds of organizations and support the ConnectWise supply chain compromise.
ServiceNow is expanding its NVIDIA partnership with Arc, a personal AI agent governed through enterprise security and access controls.
Harvard physicist Christopher Stubbs has joined Anthropic's Long-Term Benefit Trust, which can appoint a majority of the company's board.
Dario Amodei cited nine decisions that he says reduced commercial opportunities for Anthropic but improved AI safety, while accepting responsibility for weaknesses in some company policies and communications.
Several Google DeepMind employees have resigned or plan to leave over concerns about the company's military AI contracts and oversight of classified uses.
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.