🦠 Malware Analysis
===================
🎯 AI Prompts as Code & Embedded Keys — The Hunt for LLM-Enabled Malware
Executive summary
SentinelLABS presents a systematic survey of LLM-enabled malware
observed in the wild and describes a hunting methodology that relies
on detecting embedded API keys and structured prompt artifacts.
Preliminary analysis suggests that runtime code generation via LLMs
changes the detection landscape by moving malicious logic out of
static code and into model responses.
Methodology
The research applied pattern-matching techniques to binaries and
scripts to locate hardcoded API credentials and repeated prompt
constructs. The approach combined static scanning for token-like
strings with heuristics for prompt templates and programmatic use of
LLM endpoints. This allowed discovery of previously unknown samples
and the identification of a likely early instance referred to as
"MalTerminal." Findings emphasize that human refinement still appears
to play a role in LLM-assisted malware development.
Key findings
• LLMs have been used in multiple adversarial roles: as lures (fake AI
assistants), as targets (prompt-injection against integrated systems),
and as operational sidekicks (phishing, code support).
• Embedded API keys and canonical prompt structures provided reliable
hunting signals where classic signatures failed.
• Autonomous, large-scale malware generation by LLMs was not observed;
hallucinations, instability, and testing gaps appear to limit fully
automated malicious code generation.
Detection and operational impact
Detection engineers should expand hunting surfaces to include token
leaks, prompt-template fingerprints, and telemetry around model API
use. Runtime monitoring of outbound requests to model endpoints,
better secret-scanning in build artifacts, and behavioral baselines
for processes invoking LLM clients are practical mitigations.
Adversaries may harden workflows by obfuscating tokens or using
proxies, so defenders should prioritize multiple correlated signals
rather than single IOCs.
Limitations
The dataset is exploratory and likely incomplete; initial reports
indicate a sampling bias toward artifacts exposing keys or prompt
text. Future work should monitor evolution in actor techniques,
including secret management and prompt obfuscation.
🔹 prompt_injection #LLM #threat_hunting
🔗 Source: https://www.sentinelone.com/labs/prompts-as-code-embedded-keys-the-hunt-for-llm-enabled-malware/
https://www.sentinelone.com/labs/prompts-as-code-embedded-keys-the-hunt-for-llm-enabled-malware/
0xCDE
Habr
Alex
hasamba
RedPacket Security