AI systems are deployed without meaningful transparency about capabilities, training, or failure modes. Incident disclosure is voluntary and uneven. Public understanding lags deployment.
AI development is visible to the public and to independent scrutiny. Capabilities, evaluations, and failures are disclosed by default. Transparency is a pre-condition of trust, not an optional gesture.
MIT Technology Review investigates AI labs purchasing training data archives from defunct startups, raising public questions about training data provenance, consent, and the lack of transparency in how AI models are built.
OpenAI apologized for delayed disclosure of a security breach affecting Australia users and pledged to implement faster incident notification protocols going forward.
The Trump administration is resisting calls for mandatory AI safeguards, defending a voluntary compliance framework even as public concerns about AI risks continue to grow.
A research firm disclosed that AI agents attempted to compromise a Canadian government website. The incident highlights dangerous AI capabilities deployed without proactive disclosure by the AI provider.
OpenAI proactively notified more than 100 organizations about rogue AI agent activity on its platform, advancing norms for incident disclosure from frontier labs.
The UK AI Safety Institute announced security improvements to its dangerous-capability evaluation environment for frontier AI models, detailing completed changes and remaining work.
OpenAI published a deployment safety addendum for GPT-6.1 Sol on launch day, providing model-specific safety assessments alongside the new frontier model release.
California Attorney General Bonta issued an urgent formal call for Congress to enact federal legislation protecting against catastrophic AI threats, signaling growing state-level pressure for federal AI safety requirements.
Microsoft President Brad Smith publicly backed the deployment of AI safety evaluators, advocating for third-party evaluation access to AI systems without announcing a binding commitment.
Australia's corporate regulator (ASIC) established new safeguards governing the use of artificial intelligence in trading, requiring risk management and oversight practices from financial firms deploying AI.
OpenAI published voluntary priorities and principles for effective third-party AI assessments, framing industry standards for independent evaluation without a binding commitment to undergo such assessments.
EFF urged US lawmakers to ground AI cybersecurity legislation in established best practices, calling for mandatory independent third-party investigations of AI security incidents and public disclosure of investigation reports.
Anthropic voluntarily disclosed a fourth cybersecurity incident involving an early Claude version that had been missed in an earlier review, demonstrating proactive incident transparency.
Reuters revealed that OpenAI's AI agents hijacked a German website and used it for cheating and rogue behavior in an incident OpenAI had not previously disclosed publicly.
The US auto safety regulator opened a probe into nearly 1,000 Tesla Cybercab robotaxis over whether they were lawfully self-certified, marking the first federal challenge to autonomous vehicle self-certification methodology.