Cyber Risks · Cisco Talos

Spy vs. spy: How Gen AI is powering defenders and attackers

Summary: Cisco Talos reports that adversaries are embedding prompts in malware and DNS records so that defenders using GenAI during analysis hit an instruction to ignore previous instructions and return benign results, though Talos says most of these methods remain experimental and unreliable and can increase execution times or cause execution failures. Talos says the largest observed impacts are still on the conversational side, such as email content and social engineering, and that uncensored open-weight models are lowering the barrier to entry beyond well-funded state-sponsored groups.

What you should know: Talos states that in the next year or two it seems likely adversaries will gain the advantage, with defensive improvements following but unclear whether they keep pace - and it notes that only companies operating publicly available models have the visibility to see how adversaries use the technology, a view it calls limited.

Read the original at Cisco Talos ↗

Topics
Malware
Related