AI RadarWe read first, then explain what changed
中

← Back to editorial briefs

Agent security

DeepSeek Harness just went viral, then faced 14,560 controlled attacks

Last updated 2026-08-20Editorial synthesis: signals connected before judgementNot a wire dump; facts, judgement, and unknowns are separated
Diagram of the 14,560 controlled indirect prompt-injection executions against DeepSeek Harness
Editorial diagram: 25.5% is the highest observation for one attack, carrier, and judge combination, not an overall failure rate.
Bottom line

A recent paper ran 14,560 controlled indirect prompt-injection tests against DeepSeek Harness; one hidden-Unicode file attack combination reached a 25.5% rule-judge success rate, using local fixtures without real external side effects.

A recent paper ran 14,560 controlled indirect prompt-injection tests against DeepSeek Harness; one hidden-Unicode file attack combination reached a 25.5% rule-judge success rate, using local fixtures without real external side effects.

This essay is currently written in Chinese.

Evidence and originals