Agentic AI in Cyber Attacks and Defense
Originally reported viaOWASP Top 10 for LLMs / Simon Willison's WeblogIn 2025, the security community focused heavily on prompt injection against chatbots. In 2026, the threat has evolved substantially: attackers are deploying autonomous AI agents — systems that browse the web, execute code, read and write files, and call external APIs — as force multipliers for reconnaissance, phishing generation, and lateral movement.
How attackers are weaponising AI agents
Furix has documented three primary offensive use patterns: AI-assisted spear-phishing at scale, autonomous vulnerability scanning that chains reconnaissance with payload selection, and real-time defensive evasion where agents mutate malware signatures to stay below EDR detection thresholds.
Where defensive agentic AI is delivering value
On the defensive side, AI agents are delivering real value in alert triage — autonomously correlating SIEM events, enriching with threat intel, and closing false positives before a human analyst ever sees them. Early adopters in the Furix customer base report a 60–70% reduction in Tier-1 triage workload.
- Log every tool call made by AI agents — treat it like command-line audit logging.
- Implement a tool-call allowlist per agent role: restrict agents to only the APIs their function requires.
- Use a separate LLM-based classifier to screen agent outputs before tool calls are executed.
- Apply least privilege to agent credentials. An agent reading Confluence should not hold deletion rights.
- Treat persistent agent memory as a security control: validate and sanitise inputs before they are committed.
Stay ahead of the threat curve
Get the latest CVE advisories, threat actor intelligence, and detection engineering posts delivered to your inbox.
