EU tests Mythos 5, Harvey reaches $15.5B, and 3 more AI moves

In brief

• EU Cyber Access: The European Commission confirmed that ENISA received testing access to Anthropic's Mythos 5 and OpenAI's GPT-6-Astra.
• Enterprise Consolidation: Legal AI platform Harvey raised $550 million at a $15.5 billion valuation and acquired security startup Guardrails AI.
• Open World Models: Robbyant released the full model suite for LingBot-World 2.0, introducing 14B weights and a 1.3B causal-fast variant.
• State Auditing Mandate: California enacted Senate Bill 813 and Assembly Bill 1405, creating the first state statutory registry and framework for independent AI auditors.
• Research Automation: The IdeaAMBIG benchmark showed leading foundation models identify only 9.6% of implementation defects in real-world research ideas.

Five developments

1. EU Cybersecurity Agency Granted Access to Anthropic's Mythos 5 and GPT-6-Astra

• The European Commission confirmed on September 10 that the European Union Agency for Cybersecurity (ENISA) has been granted direct access to Anthropic's Mythos 5 frontier model and is testing the system for cyber-defense resilience 1.
• Commission spokespeople stated that ENISA concurrently received evaluation access to OpenAI's GPT-6-Astra model under technical protocols designed to assess infrastructure security 1.
• The deployments follow months of transatlantic discussions over sovereign oversight and provide European regulators hands-on testing environments for frontier reasoning architectures 1.

2. Legal AI Startup Harvey Raises $550M at $15.5B Valuation, Acquires Guardrails AI

• Enterprise legal AI platform Harvey closed a $550 million financing round co-led by Diffusion and Lightspeed Venture Partners, pushing the company's private valuation to $15.5 billion 2.
• Harvey concurrently announced the acquisition of San Francisco-based agent security platform Guardrails AI for an undisclosed amount, absorbing Guardrails' runtime guardrail and safety validation software into its enterprise suite 2.
• The transaction marks a steep step-up from Harvey's $11 billion valuation in March 2026, driven by contract deployment across corporate legal teams and roughly 80% of the top 100 grossing U.S. law firms 2.

3. Robbyant Releases Remaining LingBot-World 2.0 Weights, Adding 14B and 1.3B Variants

• Open-source robotics lab Robbyant shipped the final model variants of LingBot-World 2.0, providing public open weights for its interactive physical simulation architecture 3.
• The release delivers full 14B parameter causal-pretrained and bidirectional checkpoints alongside an optimized 1.3B causal-fast model tailored for lower-latency deployment 3.
• The family supports 720p rendering at 60 frames per second over unbounded simulation horizons and embeds an agentic pilot-director harness for continuous robot policy training 3.

4. California Signs Nation-First AI Auditing Framework into Law

• California Governor Gavin Newsom signed Senate Bill 813 and Assembly Bill 1405 into law on September 9, establishing the state's initial statutory framework for third-party AI system verification 4.
• SB 813 creates an authorization structure for independent verification organizations to evaluate commercial AI systems against state consumer-protection and algorithmic bias laws 4.
• Companion measure AB 1405 establishes a mandatory registry for certified AI auditors and enforces strict conflict-of-interest disclosures and operational independence standards 4.

5. IdeaAMBIG Benchmark Pinpoints Missing Method Details as Primary Bottleneck for AI Codification

• Yale University researchers released IdeaAMBIG, an empirical evaluation suite analyzing how 13 state-of-the-art foundation models handle implementation gaps across 660 research specifications 5.
• The top-performing system achieved a 9.6% recovery rate on real-world implementation defects, whereas models succeeded 80.6% of the time when explicit defects were pointed out to them 5.
• Providing precise resolutions improved downstream code implementation readiness from 14% to 98%, isolating defect localization as the central hurdle for autonomous research systems 5.

The read-through

Watch next

• European Agency Evaluation: Initial security benchmarks from ENISA's testing of Mythos 5 and GPT-6-Astra, expected in upcoming cyber resilience briefs.
• Auditor Certification Standards: The California Department of Technology's timeline for publishing official auditor registration criteria under AB 1405.
• Enterprise Security Integrations: Further consolidation among specialized runtime safety startups as vertical AI platforms seek defensible agent infrastructure.