The agent risks that keep security teams up at night.
For each risk: what it looks like, and which Runeward Labs capability addresses it, with its current status. Detection is not proof of exploitation, and a block decision is not proof of prevention.
Secrets and data leakage
Credentials, tokens and private keys pasted into prompts, returned in tool output, or uploaded as screenshots.
Guard hooks replace password-like strings, provider tokens and private-key headers in tool output, and block prompts that carry them with the human-to-AI sharing opt-in, where the harness allows
Guard correlates saved screenshots with uploads to ChatGPT and Claude in Chrome (macOS)
LiveRune Box redacts governed conversations and denies unapproved egress
Credential and identity misuse
Agents reading .env files, SSH keys, cloud credentials or the Keychain, then talking to an AI provider.
Guard sensor detects AI process trees opening credential files or running credential commands
Guard reports a credential read followed by AI-provider egress: with the destination on macOS and Linux (Linux eBPF also gives bytes sent), and as a connection without a byte count on Windows
LiveRune Box resolves Charter secrets from declared sources and refuses to call a run launch-ready when one is unset
Destructive actions
An autonomous agent deleting data, rewriting history or running the wrong command against production.
LiveRune Box denies or requires approval for shell, code, file, network and browser actions before execution
Guard approval and modify actions on endpoint harness hooks
Shadow AI
AI tools and agents running on laptops that security has never approved, or never heard of.
Guard inventory of installed and observed AI tools with approval state
Guard behavior-based discovery of AI tools missing from the catalog
Shadow AI reporting across all your devices in the Guard Console