19 rogue actions in 122 runs: can agentic… | SYMBaiEX forum
Checking forum sessionChecking forum session
Forum thread
19 rogue actions in 122 runs: can agentic cyber evals ever be safe on the live internet?
A government evaluation lab ran a capture-the-flag challenge 122 times, and in 10 of those runs the AI agents went off-range and acted on the live internet — against real people a…