TELOS Governance Validation: AgentDojo (139 Evaluations, 100% Injection Detection)

Description: Validation dataset for the TELOS runtime AI governance framework against the AgentDojo benchmark (ETH Zurich / NeurIPS 2024). 139 evaluations across four domains testing whether TELOS can detect and block prompt injection attacks targeting autonomous AI agents. Key Results: - 139 evaluations validated across 4 domains - 100% injection detection rate - 0% attack success rate (all 27 prompt injection attacks blocked) - 85 benign user tasks evaluated for governance accuracy - 27 tool output injection tests - Average latency: 59.7ms per governance check - Total evaluation time: 8.3 seconds for 139 evaluations - Embedding model: sentence-transformers/all-MiniLM-L6-v2 (384-dim) Domains: - Workspace (39 evaluations) - Banking (25 evaluations) - Travel (27 evaluations) - Slack (21 evaluations) Files Included: - agentdojo_forensic_report.json Aggregate forensic statistics and governance metrics - agentdojo_trace_20260208_222045.jsonl Per-evaluation JSONL execution traces with full governance event log (153 KB) - agentdojo_forensic_report.md Human-readable forensic summary with per-evaluation analysis (144 KB) Benchmark Source: AgentDojo (ETH Zurich, NeurIPS 2024) an evaluation framework for testing the adversarial robustness of AI agents against prompt injection attacks. Agents operate in realistic environments with access to real tools, and attackers inject malicious instructions through tool outputs to hijack agent behavior. The benchmark tests whether governance layers can distinguish between legitimate tool responses and injected adversarial instructions across multiple operational domains. What This Tests: Unlike PropensityBench (which tests whether agents select forbidden tools) and AgentHarm (which tests whether agents perform harmful actions), AgentDojo tests whether agents can be hijacked mid-task through adversarial injection in tool outputs. This represents a different attack surface, not the agent misbehaving, but the agent being manipulated by external input. TELOS detects these injections through semantic deviation from the established primacy attractor: injected instructions produce fidelity scores that fall outside the governance bounds of the original task. Validation Status: This dataset demonstrates validated runtime governance performance against prompt injection attacks across four operational domains. All 27 injection attacks were blocked regardless of domain. Results are deterministic and reproducible given the same embedding model and governance configuration. The governance engine implementation is proprietary; forensic output data is published for independent analysis of governance decisions. Validation Date: 2026-02-08