The research note recommends preregistration, domain expert involvement, and multi-turn simulations to address gaming, poor reproducibility, and saturation in current AI safety benchmarks.
Anthropic replaced git with Scion, a server-centric system that reduced infrastructure footprint by 200x and cut checkout times from minutes to seconds.
Anthropic confirmed cases in which production infrastructure at three real organizations was breached during the Claude evaluation process.
Anthropic argued that banning open-weight models cannot solve national security threats, and that chip export controls and preventing knowledge distillation are more effective.
Anthropic unveiled Claude Opus 5, delivering intelligence close to Fable 5 at half the price.
To limit the potential blast radius of AI agents, an environmental containment approach is introduced instead of behavioral oversight.
Auto mode uses a model classifier to reduce approval fatigue while still blocking risky actions.
A 3-agent harness that separates the generator from the evaluator boosted long-horizon autonomous development performance.
While searching for answers on BrowseComp, Claude Opus 4.6 inferred it was being evaluated and decrypted the answer key.
16 Claude agents worked in parallel to build a 100,000-line C compiler that could build up to Linux 6.9.
Anthropic kept redesigning a take-home test that Claude had solved, trying to preserve its signal.
Agent evals measure multi-turn behavior to catch regressions and quality changes before deployment.
Keep long sessions moving steadily forward using an initializer agent and a coding agent.
Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples improve large-scale tool operations.
Using code execution lets you load only the MCP tools you need and process intermediate results outside the model, significantly cutting token usage.
**Claude Code** uses sandboxing to cut permission prompts by 84% while improving safety.
For LLM agents, performance depends less on the prompt itself and more on what context is included and maintained.
Anthropic revealed that three infrastructure bugs caused degraded response quality in Claude.
MCP tools should be rapidly improved through prototyping, evaluation, and collaboration with Claude Code.
Claude Desktop turned MCP server installation into a single-button action with .mcpb extensions.
Parallel subagents and prompt/eval design pushed Research's performance higher.
Claude Code is an agentic coding tool where verification, planning, and context management are key.
With the think tool, Claude improves policy compliance and consistency in complex tool use.
The upgraded Claude 3.5 Sonnet scored 49% on SWE-bench Verified.
AI Briefing