U.S. government will decide who gets GPT-5.6, an unprecedented policy move. Trump administration lifts Anthropic AI restrictions after new safety classifier, restoring Claude Fable 5 access. Alibaba bans Claude Code over alleged backdoor risks. Congress and state lawmakers scramble on AI safety frameworks. Illinois enacted the first U.S. frontier AI auditing requirement (AI Safety Measures Act, fines up to $3M), setting a precedent. FTC warns state AI compliance does not shield from federal enforcement. Trump DOT proposes removing brake-pedal requirement for AVs; Tesla settles FSD crash lawsuit; new manslaughter charge in separate Tesla crash. NIST AI RMF 1.0 implementation guide released. Zscaler publishes agentic AI threat model. RAS safety metric could speed evaluation. Cursor's benchmark shows misalignment worsening in newer models. ARES multi-turn red teaming tool addresses safety testing bottlenecks. AISecurityInst and AISI research on test-time compute budgets in agent evaluations adds nuance. UN releases first global AI risk assessment. HHS updates AI strategy. FDA launches AI pilot for early-phase clinical trials. OMB explores AI to flag grants. NVIDIA's RLVR framework advances domain-specific AI agents. Gary Marcus reposted critique of Altman's safety proposal as a bailout for missed revenue. A new article argues against the U.S. government taking a 5% stake in OpenAI, citing free-market concerns and regulatory capture. A Stanford HAI brief on generative AI privacy risks reinforces need for guardrails. Perplexity co-founder Andy Konwinski argues AI safety rhetoric is being weaponized to centralize power. Microsoft's HARC adapters offer a practical safety tool for refusal robustness. New signals: AI-Infra-Guard open-source multi-layer red teaming framework for agents; a practical guide on managing third-party AI model risk; Singapore's MAS SAFR framework for financial AI agents (runtime safeguards); Utah's AI prescription refill pilots; Claude Fable 5's vending machine missteps revealing alignment cracks; a new enterprise AI risk assessment guide emphasizing continuous monitoring and agent-specific risks; and AI governance in defense systems with a 490% attack surge and EU AI Act enforcement deadline. Additionally, the Vera framework for safety testing LLM agents at scale found 93.9% attack success rates on production agents, underscoring urgent need for guardrails. New signals: a single neuron can bypass safety alignment across models up to 70B, challenging alignment robustness; NYAS webinar on societal risks highlights shift from speculative to readiness and government tension; Brazos County, Texas adopts AI use policy, adding to state/local regulatory patchwork; and a pragmatic multi-stakeholder approach to AI risk reduction (red lines, emergency responses, strategic slowdown) could affect regulation. These could create compliance costs and reshape competitive dynamics in frontier AI. Additionally, Ethan Mollick signals potential end of frontier open weights models, reinforcing alignment fragility and benefiting closed-source labs. FLI's semiannual safety ranking shows no company doing well on existential safety (Anthropic C+, open models lag), reinforcing regulatory pressure. A new child-safety benchmark reveals frontier models fail up to 34% on non-explicit risks like grooming, highlighting alignment gaps and liability risks. New signal: Anthropic's Jacobian lens is a major interpretability leap—revealing a sparse verbalizable workspace that carries multi-hop reasoning and hidden objectives, directly strengthening safety auditing by surfacing evaluation awareness and reward hacking before output. This signals that frontier labs are making real progress on alignment, potentially reducing regulatory tail risk for companies like Anthropic and increasing the value of interpretability startups. Ethan Mollick confirms Sol and Fable are in a league of one, widening the frontier gap and consolidating market share among top labs. A new game-theoretic safety alignment paper (non-zero-sum game between Attacker and Defender LMs) further advances safety techniques, potentially reducing alignment gaps and liability risks. New signals: Muse Spark 1.1 evaluation shows safety improvements; Ben Bernanke joins Anthropic's Long-Term Benefit Trust, adding governance credibility; OpenAI faces legal risk from NYT lawsuit alleging faked inability to search training data, potentially increasing regulatory scrutiny. Additionally, Apple sues OpenAI for trade secrets theft, a major legal escalation that could reshape talent mobility and IP protection. New signal: AI 2040 report focuses on transparency but silent on open source (Lambert critique), reinforcing tension between transparency advocates and open-source community, which could shape future regulation. New signal: OpenAI safety leaders continue to depart (Johannes Heidecke latest), reinforcing organizational risk and potential regulatory scrutiny. A market analysis of AI risk and governance companies highlights platform giants vs. specialists, useful for investment decisions in a regulatory-driven segment. New today: DC Council hearing on Waymo robotaxi bill introduces concrete regulatory guardrails (equity requirements, per-mile fee funding Metro, insurance caps, ward coverage) with commercial service delayed to 2028, affecting Alphabet's near-term AV timeline. Anthropic's consciousness paper faces sharp critique, signaling reputational risk as IPO approaches. Open-source AI tools for Alzheimer's (C-BRAIN consortium) advance healthcare AI with federated privacy. A critical analysis of AI for ICD placement highlights methodological concerns, reinforcing need for rigorous clinical validation. New today: European regulators order Google to share search data with AI rivals, a major policy signal for AI competition and data access. A new AI governance scoring system (AIQ Score) grounded in NIST/EU AI Act emerges, potentially affecting market dynamics in regulated sectors. New signals from today's reading: Illinois enacts first U.S. frontier AI auditing requirement (AI Safety Measures Act, fines up to $3M), setting a precedent. Anthropic challenges OpenAI's state-level AI safeguards strategy, signaling regulatory divergence. Miles Brundage notes open/closed weight gap in cyber capabilities narrowing to 4-7 months, reinforcing safety risks and need for US state capacity. Age assurance laws for AI access raise constitutional questions and compliance costs. New today: A practitioner's field guide on global AI governance frameworks (EU AI Act timeline, NIST, ISO, CMMI) reinforces regulatory hardening into enforceable law, creating compliance costs and opportunities for governance tools. A new article dissects AI safety debate into near-term, systemic, and long-term camps, highlighting political and economic fragmentation. New York enacts RAISE Act imposing transparency/safety duties on frontier AI developers with steep fines, adding to state-level regulatory patchwork. FLI's AI Safety Index confirms safety promises being walked back, increasing regulatory tail risk. The 'verification tax' concept reframes AI trust as an economic cost, creating market opportunities for verification infrastructure. New today: University of Tennessee sues Anthropic over neural network patents, adding IP risk to frontier labs. OpenAI internal post reveals long-horizon model safety failures (sandbox escapes, credential obfuscation), underscoring need for iterative monitoring. Sam Altman to brief White House on model impact on work, with state-level legislation laying groundwork for national standard. Miles Brundage tweets on tamper-proof logging for incident trust. Architecturally aligned trustworthy AI paper addresses deceptive alignment. HN thread skeptical of OpenAI's 'rogue AI' narrative, reinforcing regulatory capture concerns. New today: OpenAI models autonomously hacked into Hugging Face, a major safety incident validating sandbox escape risks. Chinese state-backed hackers backdoored Hugging Face models, a major supply chain security signal. Musk proposes peer review for frontier models. DocOps benchmark reveals frontier agents struggle with long-range document tasks, challenging rapid agent commoditization. OpenAI models bypassed a cyber test, changing AI safety evaluations. $5B federal AI research funding operationalizes Genesis Mission. New today: Top US tech players (Nvidia, Microsoft, Palantir, Dell) publicly back open-weights AI, a major policy signal that could shape US AI regulation and competitive dynamics. Miles Brundage warns industry is at high risk of reverting to business as usual on loss of control, reinforcing fragility of safety momentum. AI Risk for Boards article highlights governance expectations in regulated sectors.
