AI digest: Agent swarms and security access
Open-source agent swarms challenge closed models while governments get special AI access for cyber defence.
Big week for AI agents and who gets to use the most powerful models.
Moonshot AI drops open-weight agent swarm model
Moonshot AI released Kimi K2.6 with proper agent swarm capabilities, scaling to 300 sub-agents running 4,000 coordinated steps. The Chinese lab claims it matches GPT-5.4 and Claude Opus 4.6 on coding benchmarks whilst being completely open-weight. This feels like the first serious challenge to closed frontier models from an agent architecture perspective.
OpenAI launches cyber-specific GPT variant
OpenAI scaled its Trusted Access programme with GPT-5.4-Cyber, a fine-tuned variant built specifically for verified security defenders. Moving from pilot to broad deployment suggests they’ve worked out the vetting process for who gets access to “cyber-permissive” AI. Smart move given all the hand-wringing about dual-use capabilities.
NSA quietly using Anthropic’s most powerful model
The NSA is reportedly using Anthropic’s restricted Mythos model for intelligence work, despite ongoing tensions with the Pentagon. This is the model Anthropic deemed too dangerous for public release but apparently fine for America’s electronic surveillance agency. The optics here are fascinating given Anthropic’s safety positioning.
Anthropic secures another $5B from Amazon
Amazon pumped another $5 billion into Anthropic, with Anthropic pledging $100 billion in AWS spending in return. These circular cloud deals are becoming the new normal, but the scale suggests Anthropic is planning truly massive compute infrastructure. Reports of $30 billion annualised revenue and trillion-dollar valuations suddenly make more sense.