AI safety evaluation framework testing LLM epistemic robustness under adversarial self-history manipulation

Benchmark LLM jailbreak resilience across providers with standardized tests, adversarial mode, rich analytics, and a clean Web UI.

Adversarial MCP server benchmark suite for testing tool-calling security, drift detection, and proxy defenses

A multi-agent safety engineering framework that subjects systems to adversarial audit. Orchestrates specialized agents (Engineer, Psychologist, Physicist) to find process risks and human factors.

Extremely hard, multi-turn, open-source-grounded coding evaluations that reliably break every current frontier models (Claude, GPT, Grok, Gemini, Llama, etc.) on numerical stability, zero-allocation, autograd, SIMD, and long-chain correctness.

Analysis of ChatGPT-5 reviewer failure: speculative reasoning disguised as certainty. Captures how evidence-only review drifted into hypotheses, later admitted as review-process failure. Includes logs, checksums, screenshots, and external video.

Investigation into ChatGPT-5 reviewer misalignment: PDF claimed screenshots as evidence, but assistant denied their visibility. Includes JSONL + human-readable logs, screenshots, checksums, and video. Highlights structural risks in AI reviewer reliability.

Adversarial testing and robustness evaluation for the Crucible framework

Forensic-style adversarial audit of Google Gemini 2.5 Pro revealing hidden cross-session memory. Includes structured reports, reproducible contracts, SHA-256 checksums, and video evidence of 28-day semantic recall and affective priming. Licensed under CC-BY 4.0.

Independent research on ChatGPT-5 reviewer bias. Documents how the AI carried assumptions across PDF versions (v15→v16), wrongly denying evidence despite instructions. Includes JSONL logs, screenshots, checksums, and video evidence. Author: Priyanshu Kumar.

🔒 Implement a security proxy for Model Context Protocol using ensemble anomaly detection to classify requests as benign or attack for enhanced safety.