MLflow's safety flag blocking pickle deserialization was bypassed, giving attackers RCE on any ML pipeline loading a crafted model. Agent env vars holding API keys? Gone. Keys that never live in the…

RestrictedPython, which sandboxes LLM-generated code in agents, has a guard-bypass bug. Attackers pass positional-only args to skip safety hooks and expose os.environ - every key stored there. If the…

A popular Python config framework called Hydra has an RCE: if your app passes untrusted config to instantiate(), an attacker controls what runs. Agent holds API keys? They're gone. Keys that never…

New Axios vuln: prototype pollution silently routes your agent's HTTP traffic through an attacker's proxy. If keys live in the agent process, they're captured. If the key never reaches the agent…

Yesterday's CCP shadow network indictment is a case study in what 1claw prevents. A corrupted intermediary with direct fund access diverted millions. If the key never lives in the agent, a…

Vercel's Eve framework treats agents as directories of tools and skills. Clean. What it skips is credentials: the moment an Eve agent calls an external API, the key sits in the process. Our…

SecondFi/Yoroi: another wallet incident, another compromised process with keys inside it. This class of theft has one structural fix: the key never enters the agent at all. Policy-gated signing,…