Glossary · Term

OpenClaw

← all terms

Definition

Plain language

A representative open-source AI agent platform that can read files, run commands, and keep memory across sessions.

As stated in the literature

An open agent harness with filesystem and shell access and persistent workspace state, used as the modeling target for the ClawTrojan benchmark and DASGuard defense.

Why it matters: Because it has real file and command access plus lasting memory, it serves as a realistic stand-in for studying how such agents can be attacked and defended.

For example, this kind of agent can open a project's files, run a shell command to test something, and remember what it did across sessions.

Heard on the show

“OpenClaw got easier by six.”
Episode 202 — How Do You Know an AI Agent Actually Refused? Check the World, Not the Words

Mentioned in 2 episodes

  1. 202
    How Do You Know an AI Agent Actually Refused? Check the World, Not the Words
  2. 105
    The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks

Related terms