What changes when attackers use AI
The underlying server risks remain familiar: credential attacks, reconnaissance, exploit attempts, data theft, and service disruption. AI can make those operations cheaper to vary and faster to repeat. That makes correlation and bounded automated response more important, not less.
Server-side signals
- Rapid movement across different reconnaissance and exploit-probe categories.
- Prompt-injection and instruction-extraction probes reaching AI-backed web routes.
- Repeated activity spanning authentication, application, database, VPN, and edge logs.
- High-confidence deception paths that normal users should never request.
Vexyl Guard stores defensive classifications and redacted context. Public pages describe coverage at a useful level without publishing exact bypass conditions.
AI runtime checks
The local threat-intelligence module can evaluate application events that an authorized operator explicitly sends to it, including prompts, external or retrieved content, agent plans, tool calls, model API use, memory writes, and supply-chain changes.
- External documents and RAG content never inherit system or developer trust.
- Tool calls require task scope, user scope, and tool-policy permission.
- High-impact or irreversible actions require human approval.
- Risk decisions return a bounded score, matched defensive rules, redacted evidence, and a suggested action.
Threat families covered by the defensive model
Coverage includes prompt injection, safety bypass attempts, retrieval and memory poisoning, agent goal hijacking, tool misuse, insecure output handling, sensitive-data exposure, model extraction, poisoning, resource exhaustion, AI supply-chain compromise, malicious model-service indicators, AI-integrated malware behavior, and AI-enabled fraud indicators.
The intelligence store excludes runnable exploit code, malware code, complete jailbreak payloads, and step-by-step offensive instructions.
What Vexyl Guard does not claim
There is no reliable label that proves whether every request, script, or attack was produced by AI. Vexyl Guard focuses on observable behavior, trust boundaries, risky action context, and defensive controls. It does not market uncertain attribution as fact.