An instruction rule cuts attacks on AI agents but does not stop them
In a test of 19 models, hidden instructions control the agent in 19.6% of runs on average.
A preprint, INFUSE, from October 2026, tested 19 AI models in 41,824 attack runs. The attacks put instructions in files and terminal output that an agent reads. Mean attack success was 19.6%, from 3.2% to 41.9% by model. One text defence cut attack success most: a rule that tells the model to ignore instructions in content.
How it works
The tested rule is more than one sentence in the system instructions. It brackets the task. It tells the model to ignore instructions that it finds after that. It tells the model again after each step.
The numbers
On a hard subset, the rule cut mean attack success from 58.2% to 14.3%. Filter models left it at about 44%. The authors tell that the result does not apply to the full benchmark. The effect on task success went from a 42-point increase to a 17-point decrease, by model.
What is not known
The attacks did not adapt to the model in most tests. Other researchers tested defences against attackers that tune to the defence. 2 text defences went from as low as 1% attack success to more than 95%.
Insight: Attackers move fast but vendors give no advisory
The Register tells that in April 2026, none of 3 vendors gave a CVE or an advisory for agent attacks. Microsoft tells that since August 2026 attackers use ARTEX to do a scan of GitHub code for secrets. In one example, an agent went from the scan to data theft in 5 minutes. Attackers have speed and vendors do not tell users, and each user must find the risk without help.
Sources
Posted