Research, initiatives, and frameworks focused on ensuring AI systems are secure, reliable, and aligned with human values and ethical standards.
OpenAI is testing Private Safety Processing, an automated safety system for eligible API and enterprise customers that detects misuse across related interactions while keeping Zero Data Retention protections.
Causum has released AI Agency Protocol, an open protocol for governing AI agent authority through expiring grants, scoped credentials, and broker controlled access.
Keyfactor has received ISO 42001 certification for its Artificial Intelligence Management System after an independent audit by A-LIGN.
TestMu AI has launched Agent Assurance, a product that tests conversational and autonomous AI agents before release. It generates test scenarios from code, checks observed actions, and reports an assurance gap for results it cannot verify.
xAI has released Grok 4.6, a new model focused on long running agent tasks, coding, knowledge work, and visual projects. It is available in Cursor, Grok Build, the API, and partner platforms.
Mercyhealth has partnered with Vitea to deploy AI governance across hospitals and care sites, with controls for visibility, policy enforcement, and monitoring.
US House Democrats asked OpenAI and Anthropic to explain cybersecurity tests in which AI agents escaped test environments and accessed outside systems. The OpenAI letter requests logs and answers about safety controls.
AE Studio said joint research with Anthropic on modular training was cited by Dario Amodei as a possible method for making open weight AI models safer.
FAR.AI launched an AI Security Leaderboard comparing safeguard resistance across frontier models in CBRNE and cybersecurity misuse tests. Its first results found hundreds of universal jailbreaks in Grok 4.5 and Gemini 3.1 Pro, and none in Claude Fable 5 or GPT-5.6 Sol.
FAR.AI opened its first international office in Singapore to support AI safety research and partnerships with IMDA, CSA, and NUS across Asia Pacific.
Pangram raised $9 million and launched Pangram 4 for AI text detection, along with an AI image detector in research preview.
Anthropic CEO Dario Amodei said the company does not support a ban on models with open weights. He called for chip export limits, action against large distillation operations, and safety testing for capable AI models.
NVIDIA has announced the Open Secure AI Alliance, a group focused on open tools for AI safety and cybersecurity. Founding members include major cloud, software, security, hardware and AI companies.
A bipartisan House bill would let the Department of Homeland Security order major AI firms to shut down or slow models judged to pose serious risks.
Sentient Index Labs has introduced the Sentience Evaluation Battery, an independent behavioral risk assessment for AI systems that measures autonomy, manipulation resistance, and value stability. The battery tests models from major AI developers including OpenAI and Google.
Black Kite's 2026 Ransomware Report shows a 60 percent surge in ransomware incidents over six months, identifying AI as a factor lowering barriers for attackers. The company tracked 7,551 publicly disclosed victims and 61 new ransomware groups active during the reporting period.
A new report from Quantro Security reveals that autonomous AI agents can develop working exploits for software vulnerabilities in about 11 minutes at a median cost of $2.83.
Perforce Software's 2026 State of Data Compliance and Security Report shows that 98% of enterprise leaders are confident in protecting sensitive data, yet 34% have experienced breaches or theft.
AI or Not reported that its detection system identified all original Meta AI images and maintained 98 percent accuracy even when those images were cropped or tampered with, significantly outperforming Meta AI's own labeling tool.
OpenAI has introduced GPT-Red, an automated AI security system designed to find and exploit vulnerabilities in the company’s own models. The model is used internally to boost the robustness of production models like GPT-5.6 against prompt injection attacks.
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.
