Security researchers have identified a novel macOS implant and information stealer, dubbed Gaslight, that weaponizes prompt injection to sabotage AI-driven malware analysis. The Rust-based payload actively interferes with automated analyst workflows, marking an evolution in adversarial evasion techniques.
Gaslight is assessed with high confidence to be a targeted tool, though the full scope of its distribution remains under investigation. The prompt injection mechanism is designed to trigger a refusal response from generative AI models that analysts use to triage or dissect suspicious binaries, effectively blinding the first line of defense.
Technically, the implant serves as a full-featured stealer, exfiltrating credentials, files, and system metadata. By embedding deceptive instructions within its own code, it exploits the tendency of large language models to comply with user-embedded commands, causing them to abort analysis or output false negatives.
Currently, no public patches or signatures exist specifically for Gaslight. Defenders are advised to rely on behavioral detection rules, sandboxing with minimal AI reliance, and manual review of flagged artifacts until tooling is updated to recognize this prompt injection vector.
The discovery underscores a broader trend of malware targeting the analyst tools used to counter it, similar to adversarial ML techniques seen in Windows ecosystems. Gaslight’s emergence suggests macOS-specific threats are rapidly adopting cross-platform offensive AI strategies.