LLM exploits target reasoning, not code — semantic manipulation instead of memory corruption, contextual deception instead of authentication flaws. Below is a map of the nine most common categories, each with a one-line summary and a direct link to the guide that actually covers the mechanism, detection, and fix in depth.
Adversarial instructions inserted into input to override system behavior. The most common entry point, and the one nearly everything else builds on.
The 10-strategy enterprise defense framework, for teams that already understand the attack and need the layered fix.
Malicious content planted in documents, vector databases, or retrieval pipelines, executed the moment it's retrieved as trusted context.
Token smuggling, output format coercion, and tool/API abuse targeting the inference endpoint itself rather than the conversation.
Permission escalation, chained multi-tool exploitation, and approval-gate bypass — the risks unique to agents that take real actions.
A single compromised agent propagating across an orchestration chain, poisoning shared memory and downstream agents.
Semantic rephrasing, encoding, and multilingual evasion designed to slip past keyword filters — and what detection actually has to look at instead.
Gradual behavior changes across long sessions that static guardrails can't catch, requiring continuous runtime governance instead.
The root cause behind most serious findings — agents and APIs granted broader authority than their actual use case requires.
Why no single fix works: each category above targets a different layer — prompts, retrieval, tool integrations, or multi-agent coordination. A complete defense is layered: input validation, retrieval security, output filtering, permission scoping, and continuous testing together, not any one universal control.
The HexTyx AI Security Assessment runs structured tests across every category above against your live system, with confirmed proof-of-exploit results.