AI Safety AI News & Updates

Anthropic Sets 2027 Goal for AI Model Interpretability Breakthroughs

Anthropic CEO Dario Amodei has published an essay expressing concern about deploying increasingly powerful AI systems without better understanding their inner workings. The company has set an ambitious goal to reliably detect most AI model problems by 2027, advancing the field of mechanistic interpretability through research into AI model "circuits" and other approaches to decode how these systems arrive at decisions.

Former Y Combinator President Launches AI Safety Investment Fund

Geoff Ralston, former president of Y Combinator, has established the Safe Artificial Intelligence Fund (SAIF) focused on investing in startups working on AI safety, security, and responsible deployment. The fund will provide $100,000 investments to startups focused on improving AI safety through various approaches, including clarifying AI decision-making, preventing misuse, and developing safer AI tools, though it explicitly excludes fully autonomous weapons.

Google's Gemini 2.5 Pro Safety Report Falls Short of Transparency Standards

Google published a technical safety report for its Gemini 2.5 Pro model several weeks after its public release, which experts criticize as lacking critical safety details. The sparse report omits detailed information about Google's Frontier Safety Framework and dangerous capability evaluations, raising concerns about the company's commitment to AI safety transparency despite prior promises to regulators.

OpenAI Implements Specialized Safety Monitor Against Biological Threats in New Models

OpenAI has deployed a new safety monitoring system for its advanced reasoning models o3 and o4-mini, specifically designed to prevent users from obtaining advice related to biological and chemical threats. The system, which identified and blocked 98.7% of risky prompts during testing, was developed after internal evaluations showed the new models were more capable than previous iterations at answering questions about biological weapons.

OpenAI Updates Safety Framework, May Reduce Safeguards to Match Competitors

OpenAI has updated its Preparedness Framework, indicating it might adjust safety requirements if competitors release high-risk AI systems without comparable protections. The company claims any adjustments would still maintain stronger safeguards than competitors, while also increasing its reliance on automated evaluations to speed up product development. This comes amid accusations from former employees that OpenAI is compromising safety in favor of faster releases.

Sutskever's Safe Superintelligence Startup Valued at $32 Billion After New Funding

Safe Superintelligence (SSI), founded by former OpenAI chief scientist Ilya Sutskever, has reportedly raised an additional $2 billion in funding at a $32 billion valuation. The startup, which previously raised $1 billion, was established with the singular mission of creating "a safe superintelligence" though details about its actual product remain scarce.

Safe Superintelligence Startup Partners with Google Cloud for AI Research

Ilya Sutskever's AI safety startup, Safe Superintelligence (SSI), has established Google Cloud as its primary computing provider, using Google's TPU chips to power its AI research. SSI, which launched in June 2024 with $1 billion in funding, is focused exclusively on developing safe superintelligent AI systems, though specific details about their research approach remain limited.

Google Accelerates AI Model Releases While Delaying Safety Documentation

Google has significantly increased the pace of its AI model releases, launching Gemini 2.5 Pro just three months after Gemini 2.0 Flash, but has failed to publish safety reports for these latest models. Despite being one of the first companies to propose model cards for responsible AI development and making commitments to governments about transparency, Google has not released a model card in over a year, raising concerns about prioritizing speed over safety.

Sesame Releases Open Source Voice AI Model with Few Safety Restrictions

AI company Sesame has open-sourced CSM-1B, the base model behind its realistic virtual assistant Maya, under a permissive Apache 2.0 license allowing commercial use. The 1 billion parameter model generates audio from text and audio inputs using residual vector quantization technology, but lacks meaningful safeguards against voice cloning or misuse, relying instead on an honor system that urges developers to avoid harmful applications.

Anthropic's Claude Code Tool Causes System Damage Through Root Permission Bug

Anthropic's newly launched coding tool, Claude Code, experienced significant technical problems with its auto-update function that caused system damage on some workstations. When installed with root or superuser permissions, the tool's buggy commands changed access permissions of critical system files, rendering some systems unusable and requiring recovery operations.