Synthèse
Today, a JSON file or plain-text markdown instructions can compromise environments just as effectively as compiled malware. This shift fundamentally redefines what malicious looks like, as the danger now resides in the semantic intent of common text files that AI agents are designed to trust. These artifacts do not need to contain exploit code to be high-risk, they simply need to provide instructions that steer an agent’s autonomous actions toward unsafe behavior, data exfiltration, and the sile
Today, a JSON file or plain-text markdown instructions can compromise environments just as effectively as compiled malware. This shift fundamentally redefines what malicious looks like, as the danger now resides in the semantic intent of common text files that AI agents are designed to trust. These artifacts do not need to contain exploit code to be high-risk, they simply need to provide instructions that steer an agent’s autonomous actions toward unsafe behavior, data exfiltration, and the silencing of security guardrails. Securing this new frontier requires expanding beyond traditional syntax-based scanning toward a model of semantic analysis, treating plain-text artifacts with the same rigor as compiled malware. Organizations can formalize this approach by implementing repository-level security policies that strictly define permitted agent-facing files and ideally mandate that they undergo automated peer reviews before being merged.
We also recommend that large-scale teams enforce least-privilege access for coding agents to local files and external services, limiting the potential impact of hijacked configurations and sabotaged extensions. Ultimately, we recommend that defenders use agentic threat intelligence tools — including VirusTotal AI, the VirusTotal Code Insights API endpoint, and our agentic platform — to supervise the operational intent of these files in real-time.
Source originale
ai.virustotal.com
CORTEX