AI
Claude Sonnet 5's system card reveals the real test of reliable AI agents
Anthropic's detailed safety evaluation, not its benchmarks, exposes what truly matters for autonomous AI: handling real-world failures, stale data, and adversarial attacks.
The New Stack
Anthropic released Claude Sonnet 5 this week with the usual benchmark scorecard — higher marks for coding, reasoning, and agent tasks. But the 145-page system card accompanying the model tells a different, more revealing story about where AI engineering actually needs to improve.
Beyond benchmarks: the real safety questions#
While performance metrics occupy just a small portion of the system card, the bulk of it examines how agents behave when operating with minimal human oversight. Anthropic evaluated whether agents can browse the web without being hijacked, use tools reliably, plan across multi-step tasks, resist prompt injection attempts, and recover gracefully when something breaks mid-workflow.
The card introduces evaluations that earlier language models barely considered. It includes testing for covert agent behavior using SHADE-Arena and LinuxArena — checking whether agents pursue hidden objectives while appearing to follow instructions. Sonnet 5's stealth scores approached zero, reflecting the company's focus on transparency during autonomous operation. Similarly, Anthropic deployed a live bug bounty program that pitted adaptive attackers against its coding, computer use, and browser use environments.
The infrastructure challenge ahead#
What emerges from these assessments is less about model capability and more about the engineering patterns teams need to build around agents. As autonomous systems take on sustained tasks — investigating production incidents, reviewing code, managing documentation — they'll face practical breakdowns that no benchmark captures.
A long-running workflow can fail in many ways: a tool request times out mid-execution, a browser session loses state after a redirect, an external API goes temporarily offline. Each disruption forces the agent to recognize what changed, preserve its progress, and decide whether to retry or halt. These scenarios demand infrastructure components like state persistence across multiple steps, synchronized external tools, and failure detection before stale data propagates.
Anthropic's own evaluation framework hints at how this looks in practice — features such as tool result clearing (removing outdated outputs as context accumulates) and memory mechanisms that store information outside the active context window. These aren't flashy capabilities, but they're essential for agents operating over hours rather than single conversation turns.
