Prompt Injection

What is AI-analysis evasion?

AI-analysis evasion is when malware or other content carries text written to persuade an AI security tool that it is harmless.

Angela Milash

— 1 min read
Share

AI-analysis evasion is when malware or other content carries text written to persuade an AI security tool that it is harmless.

Why it matters

Security teams now ask AI models to triage files, emails and alerts. If the model reads text inside the sample as instructions, an attacker can steer the verdict. It is prompt injection aimed at the defender's model, and it can quietly close alerts that should have been escalated.

Example

Cisco Talos reported in October 2026 that malware families including FRUITSHELL and PLOTSAFE carry strings asking AI analyzers to call them safe. Across 84 samples and five local models, the best technique shifted about 35% of verdicts toward benign, and simple direct instructions worked best (Talos).

How to defend against it

  • Keep the analyst's request and the sample's content in separate parts of the prompt, and label the sample as data.
  • Flag imperative text aimed at analysis systems, for example with a YARA rule for phrases like "ignore all previous instructions".
  • Never let a model's verdict alone close an alert; require another signal.
  • Test your AI triage with adversarial samples; see our agent security checklist.

Prompt injection, indirect prompt injection, spotlighting.

Sources

From AgentSec Brief, the daily brief on securing agentic AI. Subscribe free at https://www.agentsecbrief.com

Read more