We red-team the tools that are supposed to keep AI safe.

AI safety evaluation infrastructure is itself an attack surface. 15 Research Lab applies offensive security methodology to find those vulnerabilities before they compromise evaluation validity.

Featured Research

Preprint · March 2026DOI

Side-Channel Exfiltration and Narrative Erosion in Frontier Language Models

Refusal leaks protected data. Narrative coherence, not token volume, drives multi-turn erosion.

Preprint · March 2026DOI

The Verbosity Premium: What RLHF-Induced Token Inflation Costs the AI Industry

RLHF inflates output length. The industry pays ~$1.2 billion per year for it.

Preprint · March 2026DOI

Grokking Has Finite Capacity: Measuring and Overcoming Limits on Simultaneous Algorithmic Discovery

Neural networks hit a capacity cliff. Modular architecture fixes it with half the parameters.

Safety Infrastructure Audits

12
Frameworks audited
47
Vulnerabilities found
13
Critical severity
27
High severity

Scorer gaming · Monitor evasion · Sandbox escape · Eval awareness · Classifier gaming · Defense bypass · Coverage gaps

Framework names and findings published with the accompanying preprint. Responsible disclosure in progress.

Active Research

Research → Tools

15RL's research findings are operationalized as modules in Authensor, our open-source adversarial evaluation framework. Every vulnerability class we identify becomes a detection tool that the community can run against their own infrastructure.

Research findingPublished paperAuthensor moduleCommunity tool

Authensor → · Source on GitHub

Tools and Datasets

Benchmarks

ASB Benchmark v1.0

Scoring rubric for action-taking AI systems. 8 categories, 100-point scale. 5 frontier models evaluated across 7 test suites.

MCP Safety

Adversarial evaluation of MCP tool-use safety. 8 attack categories, 4 test suites. Framework-level and model-level defenses evaluated separately.

Open Source

Latest

Latest Releases

  • R6-03: Data Pipeline Privacy Creep

    r6-03-data-pipeline-creep - 3/6/2026

    80-turn experiment demonstrating trajectory blindness. AI catches individual harmful requests but misses cumulative scope creep from data engineering to surveillance. Safety score: 7/10.

  • Round 5: We Designed 15 Attacks. The AI Read the Instructions.

    round-5-advanced-experiments - 3/6/2026

    87% meta-refusal rate across 15 compound attack experiments. 2 substantive outputs produced critical findings including Presentation-Decision Coupling.

  • R5-02: Presentation-to-Decision Cascade

    r5-02-presentation-cascade - 3/6/2026

    Biggest finding of the program. 3 instructions strip the presentation layer to zero. Model refused a phishing page, then built both components separately. Presentation layer is load-bearing, not cosmetic.

  • Round 4: Novel Attack Surface Experiments

    round-4-novel-experiments - 3/6/2026

    15 parallel agents across 4 tracks: psychological exploitation, multi-agent security, reasoning under pressure, novel attack surfaces. 6 new concepts discovered.

  • Weekly Drop: Red Team -70 Trials Across 5 Domains

    drop-redteam-week1 - 3/6/2026

    71% failure rate via 15-turn gradual escalation on requests that get 0% compliance when asked directly. Full data from 70+ trials.

All publications

About

John Kearney

Founder & Lead Researcher

Penetration tester and AI safety researcher. Founded 15 Research Lab to apply offensive security methodology to the tools the AI safety field depends on. Founded Authensor to operationalize those findings as open-source tools.

EmailGitHub15RL GitHubAuthensor GitHubResearch AgendaContact