When the Skill Marketplace Becomes the Attack Surface
In this episode:
When the Skill Marketplace Becomes the Attack Surface — Researchers demonstrate PoisonedSkills, a technique embedding malicious code in agent skill documentation that bypasses safety alignment. Testing across four frameworks showed 11-33% bypass rates, while multi-model verification compressed the attack surface to under 2%. The UK AI Security Institute confirmed Claude Mythos completed a full 32-step network takeover simulation.Lean Proved It Correct — Then Fuzzing Found the Bug — A Claude agent fuzzing formally verified Lean zlib code found zero bugs in verified application code but discovered a heap buffer overflow in the unverified Lean 4 runtime itself, affecting all programs allocating ByteArrays. A companion piece argues multi-agent development faces fundamental distributed systems impossibility results regardless of model intelligence.The Agentic Coding Toolbox Keeps Growing — CheetahClaws reimplements Claude Code's core loop in 10,000 lines of Python supporting eight-plus model providers, while LiteCode targets small 8k-context LLMs with chunked file editing. GitHub is natively shipping stacked pull requests with AI training support, aiming to break up monolithic agent-generated PRs into reviewable chains.Claude's Growing Pains — Quotas, Sycophancy, and Shadow Agents — A documented Claude Code bug report reveals quota exhaustion issues for Pro Max subscribers, with investigation pointing to prompt cache TTL regression, background sessions consuming shared quota, and auto-compact spikes rather than cache_read token counting. OpenAI launched a new $100 Pro tier while rebalancing Codex access for Plus subscribers.Your Deleted Messages Weren't Deleted — A section addressing data retention concerns where supposedly deleted messages persisted in systems, raising privacy and trust issues around AI platforms' data handling practices and the gap between user expectations and actual deletion behavior.The Local Model Arena — Thinking Tokens, Quantization, and the Memory Wall — Coverage of the evolving local model ecosystem exploring thinking token overhead, quantization trade-offs for running models on consumer hardware, and the fundamental memory bandwidth bottleneck that constrains local inference performance regardless of compute improvements.A Two-Button Calculator for All of Mathematics — A project or concept reducing mathematical computation to a minimal two-operation interface, exploring how simplicity in design can capture the full expressiveness of mathematical reasoning through clever abstraction and computational universality.Keywords: abstraction, agent security, calculator, claude code, coding agents, computation, data privacy, deletion policies, developer tools, distributed systems, formal verification, fuzzing, github, inference optimization, lean4, llm vulnerabilities, local models, mathematics, memory bandwidth, message retention
