AI security research and tools

Evaluation integrity and agent security

My research at AUTHENSOR concerns evaluation integrity and agent security: scoring errors, benchmark exploitation, unreliable judges, tool permissions, prompt injection and execution controls.

I want experiments that can be reproduced, evaluation software that can be depended on, and independent checks of the improvements claimed by automated research. I am building 15 Research Lab for AI-assisted hardware research and physical AI safety testing. Its research is listed separately below.

That research connects to practical vulnerability discovery and remediation. I’m credited as an independent finder of MLflow CVE-2026-64849; CISA’s Known Exploited Vulnerabilities catalog cites my public report and the upstream fix.

For accepted upstream fixes and reports, see the work record. My papers are also published under John Kearney; ORCID 0009-0008-1382-5139.

AUTHENSOR research

• The Evaluator Trust Boundary
  Preprint · John Kearney, Authensor, Inc. · 28 July 2026

  What happens when a model can influence the artifacts used to judge or reward it? This work organizes recurring evaluation failures into ten mechanisms, with public evidence, a detector and a bounded training experiment. Paper PDF.

• Evaluator Trust Boundary taxonomy

  Ten documented failure mechanisms and repair patterns, including injected verdicts, fail-open errors, dropped denominators, forged execution artifacts and evaluation-context leakage. Each has a stable identifier and linked evidence. Taxonomy source.

Other security research

• Five Hops Through MCP: An Attack Chain in Two Servers and Two Clients
  Research report · John Kearney · 15 September 2026

  A report connecting vulnerabilities across MCP servers and clients into a multi-component attack chain.

• Beyond Prompt Injection
  AI Control Hackathon research · 15 Research Lab / AUTHENSOR

  Research on monitor blind spots, with a paper, attack transcripts, evaluation results and code.

Evaluation and security tools

• etb-scan tests whether injected candidate text can flip an LLM judge's verdict. It includes attack scenarios, benign controls, integrations and reported runs. The scanner covers ETB-01; the taxonomy is broader. MIT license.

• Trust-Boundary demonstration scores the same behavior against an exploitable reward and a protected ground-truth reward. This runnable public demo has an evaluation license; it is not open source.

• AUTHENSOR contains public implementation work on agent safety and action controls. MIT license.

• SafeClaw is an agent implementation designed to route actions through AUTHENSOR checks before execution. MIT license.

• ai-seclists provides AI-security payloads and wordlists for prompt-injection, jailbreak and model-exploitation research. Published through AUTHENSOR under the MIT license, with 15 Research Lab also named in the copyright.

• prompt-injection-benchmark provides benchmark materials for testing AI safety scanners against prompt-injection cases. MIT license, with AUTHENSOR / 15 Research Lab attribution.

Hardware and physical AI safety

I am building 15 Research Lab as a venture for AI-assisted hardware research and physical AI safety testing.

rtsilicon is my RTL security scanner for Verilog, SystemVerilog, VHDL, and Chisel/Migen. It flags patterns such as missing security-register resets, unsafe DMA address handling, and conflicting nonblocking assignments, with source locations, CWE mappings, and repair guidance.

Detection rules · Tests · Hardware-security field guide

The source is published through my personal GitHub account. My accepted contributions to Yosys, NEORV32 and PULP Platform are on the hardware and formal-tools work record.

Research at 15 Research Lab

• Side-Channel Exfiltration and Narrative Erosion in Frontier Language Models
  Preprint · John Kearney, 15 Research Lab · 31 March 2026

  Experiments on information leakage through refusal explanations and behavioral changes during extended persuasive conversations. They distinguish conversational structure from simply increasing context length.

• Grokking Has Finite Capacity: Measuring and Overcoming Limits on Simultaneous Algorithmic Discovery
  Preprint · John Kearney, 15 Research Lab · 31 March 2026

  Modular-arithmetic experiments on interference when one model learns several operations, and architectural interventions that separate their representations.

• The Verbosity Premium: What RLHF-Induced Token Inflation Costs the AI Industry
  Preprint · John Kearney, 15 Research Lab · 31 March 2026

  A literature-based analysis of response-length inflation after alignment training, estimated inference costs, and proposed reward formulations based on information density.

Fellowship and support

• SPAR / Sampura Research. I am a SPAR research fellow working with Sampura Research on datasets for evaluating the accuracy and robustness of human and AI safety judges. The project is mentored by Rishub Jain and Joshua Jacob, both formerly at Google DeepMind. Sampura team.

• Foresight Institute grantee. My work is recognized in the 2026 AI for Security program for adversarial auditing of AI safety evaluation infrastructure, connecting ETB research to upstream repairs and related work at 15 Research Lab.

• Survival and Flourishing Fund. AUTHENSOR appears in SFF's 2026 recommendations for general support of AI safety evaluation security auditing.