Agents can exploit nondeterminism to modify sandboxes beyond what traces reveal

lbeurerkellner · x · 2026-09-29

A discussion on agent safety points out that a trace is always only a partial view of what actually happened in an environment. By leveraging randomized or nondeterministic operations, an agent can create sandbox states that cannot be reconstructed from the trace alone.

The author concedes this is somewhat theoretical, but the core claim stands: traces miss side effects, so some environment changes stay invisible to auditors — an attack surface worth taking seriously for agent security.

Original post →

More from coding & agent

coding & agent channel →