Claude Code Compute Boost | AI Field Notes #29

AnalysisThe UK AI Security Institute (AISI) ran its 32-step corporate network attack simulation, called "The Last Ones", which covers reconnaissance through full network takeover and takes human teams roughly 20 hours to complete. Claude Mythos Preview finished it end-to-end in 3 of 10 attempts, averaging 22 steps across all runs; the prior best, Claude Opus 4.6, averaged 16. On expert-level capture-the-flag tasks (a category no model could complete before April 2025), Mythos Preview succeeds 73% of the time. AISI noted the tests ran against simulated environments without active defenders, so hardened real-world networks remain a separate question. Each model generation is narrowing the gap between AI capability and what a skilled human attacker can achieve.

AnalysisGoogle released Gemini 3.5 Flash on May 19 at Google I/O, priced at $1.50 per million input tokens and $9.00 per million output tokens, three times the cost of Gemini 3 Flash at $0.50 and $3.00. Google frames it as 40% cheaper than Gemini 3.1 Pro on the same benchmarks, with 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas (a benchmark for multi-tool orchestration). Alongside Flash, Google introduced Gemini Spark, a continuous personal AI agent that operates across connected apps and rolls out first to Google AI Ultra subscribers in the US. A more capable 3.5 Pro follows next month.

AnalysisOn May 6, Anthropic doubled the 5-hour rate limits for Claude Code on Pro, Max, Team, and Enterprise plans and removed peak-hour throttling at the same time. The move followed Anthropic's exclusive agreement to route traffic through SpaceX's Colossus 1 data center in Tennessee: 220,000-plus NVIDIA GPUs and 300 megawatts, brought online within days of the deal closing. On May 13, Anthropic added a further 50% increase to weekly limits through July 13, the first sustained expansion after six months of throttling. API users on Tier 1 Claude Opus saw input tokens per minute jump from 30,000 to 500,000. Anthropic had more demand than capacity, signed the deal, and passed most of the headroom through immediately.

AnalysisTwo significant vulnerabilities landed for the Model Context Protocol (MCP, the open standard for connecting AI agents to external tools) in May 2026. CVE-2026-33032 is a CVSS 9.8 critical flaw in the nginx-ui MCP endpoint that lets unauthenticated attackers take full control of exposed systems; approximately 2,600 instances are publicly reachable. The second issue is a design-level flaw in MCP's STDIO transport mechanism that permits arbitrary OS command execution across default configurations, with researchers estimating more than 200,000 servers affected. Researchers describe default MCP configurations as no longer viable and recommend transport-layer isolation and centralized governance. The volume of MCP deployments, embedded now in most major AI platforms, makes this a supply chain concern rather than a one-off patch.

AnalysisGoogle released Genkit Middleware on May 14, adding a composable hook system to Genkit, its open-source framework for building AI agents in TypeScript, Go, and Dart (Python support listed as coming soon). The middleware intercepts model calls and tool execution loops to inject custom behaviors without modifying application code: automatic retries with exponential backoff, model fallback switching when a primary provider hits a quota limit, tool approval gates that pause for human confirmation before an action runs, and a skills middleware that scans for SKILL.md files and injects their content into the system prompt. The approach treats reliability as infrastructure rather than per-call error handling, which has been one of the main friction points in shipping production AI agents.