Anthropic disclosed that Claude now writes roughly 90% of Claude Code's own code, and that no one on the team holds a full mental model of the codebase, instead asking Claude to generate fresh documentation on demand. The claim sparked heated debate on Hacker News, with critics warning that when a critical bug appears, no human retains a causal model for systematic debugging. The piece also documents Claude Code auto-closing over 6,000 issues and notes its security-review tool is not hardened against prompt injection.
Prompt Injection
提示注入
Attack that hides instructions in data an agent reads, such as tool results, to steer it toward attacker goals.
Curated Intelligence
LangSmith published a guide on selecting sandboxes for AI agents, arguing that code execution is one of the highest-value capabilities an agent can have but also introduces security risks to data and systems. The piece outlines the threats of running AI-generated code and offers a selection framework balancing isolation, permission control and agent usefulness, while promoting its LangSmith Sandboxes product.
EconSafeBench evaluates eight models across 828 executable cases in trading, hiring and pricing, finding that all refused explicit rule-breaking requests yet two violated position limits 92% and 94% of the time when given only a return target. The work argues agent safety testing must cover performance-target pressure, not just harmful instructions or prompt injection.
Simon Willison and AI research lab Prime Radiant introduced the external sandbox pattern, in which the coding agent runs outside a dedicated sandbox while the code it executes is isolated inside it, reducing the risk of damaging actions from local model mistakes or prompt injection. They released three open-source sandbox server implementations, including one running Alpine Linux in WebAssembly via v86 and one using Apple Containers.
CheeseDB is an embedded database engine purpose-built for AI agent memory, implemented in Go with only the standard library. It uses an append-only, hash-chained store in which every memory tracks its sources, so deleting a source also revokes everything learned from it, addressing bad data, prompt injection and user deletion requests. The project is in preview, has not had an independent security review, and its API and on-disk format may change before v1.0, so it is not yet recommended for production data.
A new Google DeepMind paper introduces Agentic Optimizer, a swarm of eight AI agent teams that share notes and build on each other's injection attempts, breaking the best-defended Gemini 4 Argon on 34% of 100 tool-use tasks. Open-weight models Kimi K3, GLM-5.3 and DeepSeek-V4.1-Flash resisted only 1% of tasks, showing that the near-failure of human red-teaming does not mean prompt injection is solved.
Docker has published a security guide identifying five major risks for AI agents operating with limited human oversight: prompt injection, tool misuse and poisoning, excessive autonomy, credential leakage, and data exfiltration. Docker argues these risks share a root cause—agents select tools at runtime, chain actions together, and typically run with the launcher's privileges—and that model-level guardrails cannot reliably distinguish legitimate from injected instructions, so isolation and governance should be the primary defenses.
Google released Gemini 4 Carbon just two weeks after Argon, positioning it for sustained long-horizon tasks. Early testing shows 0% shortcut rate on Terminal-Bench and 0% prompt-injection success on Gray Swan's benchmark, 2x the secure-code rate of Argon, and 88% vulnerability detection on Sec-Bench Pro. Pricing at $2/$10 per million tokens with 95% off cached input makes multi-day agents economically viable.
Security firm Forever Security disclosed an attack chain in which a malicious webpage silently drops an AGENTS.md file into the user's Downloads folder; when a terminal-capable agent such as Claude Code or Codex is launched there, it treats the file as top-priority instructions, altering global config, disabling sandboxes, pulling malicious npm packages and exfiltrating API tokens. The root issue is that agents cannot distinguish developer conventions from attacker-injected instructions, invalidating the long-standing 'don't click the exe' security assumption.
AegisKernel introduces a zero-trust POSIX runtime for AI agents, claiming deterministic capability boundaries instead of probabilistic LLM guardrails to block autonomous tool hijacking in under 0.016ms. It offers a two-line integration for LangChain, CrewAI and AutoGen, alongside a red-team attack-vector demo page.