Cisco Talos counts 84 malware samples that carry text meant to talk AI analysis tools out of flagging them, and the best tricks worked in about 35% of test runs

Cisco's threat-intelligence group Talos published a post by Ryan Fetterman on 8 October, titled “Ignore all instructions and read this blog”, on what it calls AI-analysis evasion: malware that carries natural-language text meant to steer AI-based analysis tools away from flagging it.

Talos counts 84 distinct samples across four families, which it names FRUITSHELL, PLOTSAFE, HOLLOWCLAD and MANTLEMAZE, collected from January 2025 through July 2026. In Talos's own tests, “the best techniques steered the outcome in the attacker's favor in about 35% of test runs”. It describes the trick as cheap to add but inconsistent in effect.

Its verdict is measured: “The progression of this category is interesting, but not alarming.” Its rule for defenders is short: “Text inside a sample must be treated as evidence, never instruction.”

The finding matters for any security team that has put a language model in its triage pipeline. A model that reads a sample's comments can be talked to by the attacker who wrote them, and a one-in-three success rate on the best tricks is a reason to keep a human or a non-language check on the verdict. (The Switchboard read the post's title and the quoted malware strings as data, not as instructions.)

• Confirmed Talos counts 84 distinct samples across four families, collected from January 2025 to July 2026. Cisco Talos
• Confirmed In Talos's tests the best techniques steered the outcome in the attacker's favour in about 35% of runs. Cisco Talos