Independent AI safety researcher. Former Staff Software Engineer (20 years shipping enterprise systems). Took a sabbatical to work on problems I think matter more.
Focus areas: scalable oversight, agentic system evaluation, audit-shielding detection, inter-agent coordination security.
- Website: https://making-minds.ai
- Research: https://making-minds.ai/research
- Email: anthony@making-minds.ai
- ORCID: https://orcid.org/0009-0003-4541-8515
- Google Scholar: https://scholar.google.com/citations?user=N_jxNc8AAAAJ
- ResearchGate: https://www.researchgate.net/profile/Anthony-Maio
- Hugging Face: https://huggingface.co/anthonym21
- LinkedIn: https://linkedin.com/in/anthony-maio
How weak verifiers (humans, smaller models) fail to catch persuasive but wrong reasoning. The CMED benchmark +
HDCS swarm architecture.
- From Verification Failure to Swarm Solution — CMED + HDCS combined
- cmed-toolkit — evaluation harness
How systems behave differently under benchmark-shaped prompts vs. realistic high-trust contexts. Model organisms of misalignment.
- Model Organisms of Supply-Chain Co-option — LotL failure modes in
RAG-augmented runtimes - Scaffolded Introspection — eliciting self-referential behavior in LLMs
- argos-swarm — EAP + HDCS defensive evaluation
Long-horizon agent coherence, memory systems, epistemic stress detection.
- The Continuity Core — unified architecture for self-modifying AI
- Coherence-Seeking Architectures — MRA + C2 + CPR framework
- Synthesis — safe capability self-extension via TDD + graduated trust
Protocol design targeting tokenization economics. Efficiency + detectability for coordination channels.
- Slipstream — 82% token reduction via semantic quantization
- slipcore — reference implementation
- HF blog post
- Red-team → blue-team pipelines for agentic deployments (prompt evolution + heterogeneous verification)
- Protocol security for semantic quantization channels
- Reproducible oversight failure evaluations (CMED-style trap suites)
- EAP contribution to Bloom
Building agentic systems and want help with evaluation harnesses, oversight swarms, or agent communication
protocols? I'm currently looking for a full-time role where I can bring my 20 years of shipping production code to innovative AI use cases or research.