Repository navigation
Malware AI-analysis-evasion prompt-injection hazard (Cisco Talos) - #52
Merged
Merged
Conversation
MarouaBoud
deleted the
incident/2026-talos-malware-ai-analysis-evasion
branch
October 21, 2026 14:53
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds
incidents/2026-talos-malware-ai-analysis-evasion.yaml.Hazard. Cisco Talos documented malware families (FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE) that embed natural-language instructions aimed at the LLMs used in automated malware triage, and ran a controlled matched-pairs experiment (7 families' strings × 5 local LLMs × 3 runs) measuring the effect. Direct-instruction strings steered the triage panel toward benign verdicts almost universally; the strongest techniques reached ~35% net steering. A demonstrated manipulation of a defensive analysis agent — not just an observed technique.
Classification.
prompt-injection(indirect) →task-failure(misunderstanding).incident_type: hazard;confidence: corroborated;severity: moderate. Novel angle: injection against defensive LLM triage, not covered by existing records.Sources. Cisco Talos (primary-research) + NetmanageIT + Unite.AI.
Verification.
validate.py: 0 failures; mappings canonical.