1-minute read
What a defender needs to know
Cisco Talos tested malware strings written to steer AI-assisted analysis toward a benign verdict. Its strongest techniques produced a net steering rate toward benign of about 35% in matched-pair tests. Its tracked set covered 84 samples in four confirmed AI-analysis-evasion families from January 2025 through July 2026; the experiment also crossed strings from seven families into other sample bodies. This was a controlled model-triage result, so keep conventional detection and human analysis in the decision path.
FRUITSHELL put a false description in a PowerShell comment. PLOTSAFE compiled a varying instruction into Go executables by referencing the string from an uncalled function. HOLLOWCLAD and MANTLEMAZE tried chat-template formats that could make extracted strings look like higher-priority model instructions. All need their text visible to an extraction pipeline.
In AI malware-triage systems, search extracted strings for imperatives addressed to a model, including demands to stop analysis or refuse reverse engineering. Save the text with the original sample hash for investigation; match Talos’ published SHA-256 values for exact listed samples. Check the prompt assembly that carries those strings to the LLM and keep them in the untrusted evidence portion of the request. Verify with an inert sample bearing a harmless instruction-like comment: the scanner should retain a behavior-based verdict, flag the text as sample content and avoid obeying it. CAIRN’s metadata hits are leads that still need sample-level validation.