01 / Stage
Research
Formal results and empirical studies — impossibility and trade-off theorems such as the Defense Trilemma, agent-security work, AI evaluation methodology, and systems security.
[ lab // artifacts // disclosures ]
Not just papers.
Shrewd Research is a small independent research institute and security lab. The same method — research, test, falsify, build, disclose — is applied to AI wherever it is deployed: healthcare agents, multi-agent teams, model evaluation, digital forensics, and the infrastructure software AI runs on. The work runs end to end: formal and empirical research, independent verification of other people's claims, released artifacts, real vulnerability findings in production software, and a public record of where the results get used.
[ method // 06 stages ]
Research → falsification → artifacts → findings → adoption
02 / Stage
Falsification
Take a widely held assumption about a benchmark, defense, or architecture and determine its actual scope conditions. A negative or boundary result is a finished result, not a failed project.
03 / Stage
Measurement
Build adversarial benchmarks and experiment designs — Manifold of Failure, Go-Explore red teaming — that find structured weaknesses and test whether reported improvements generalize.
04 / Stage
Artifacts
Papers become usable technology: trained models, quantized builds that run locally, evaluation harnesses. Distributed openly rather than kept internal.
05 / Stage
Findings
Vulnerability research and coordinated disclosure against real software, including assigned CVEs in widely deployed open-source infrastructure.
06 / Stage
Adoption
Track where the concepts are independently invoked, implemented, or built on — kept separate from the papers themselves in the impact ledger.
[ disclosures // assigned_cves ]
Vulnerability research
The work is not confined to model evaluation. We also find and report flaws in the production software AI systems are built on, through coordinated disclosure with the maintainers.
CVE-2025-489132025
Apache CXF — unsafe JMS configuration enabling remote code execution
A JMS configuration flaw in Apache CXF that could allow an attacker to trigger RMI- or LDAP-based remote code execution through untrusted protocol handlers.
Credit
Blake Gatto (Shrewd Research) and Manish Bhatt, credited as finders.
Outcome
Apache restricted the affected protocols in the fixed releases.
[ artifacts // open_distribution ]
Released technology
Where a result can be turned into something defenders run, we build and publish it. PhishSense is the clearest case: a paper, a trained model, a locally runnable quantized build, and an evaluation story — not a conceptual proposal.
Formal proofs / Lean 4
Coffee Law — proofs & checks
The knowledge budget for agent collectives, released with machine-checked proofs and scripts that reproduce every number in the paper.
Core steps proved in Lean 4 with no unfinished proofs; five Python suites verify the worked examples and randomized tests of the general inequalities. MIT licensed.
GitHubModel / Transformers
Llama-Phishsense-1B
A domain-specific phishing detection model, fine-tuned and released publicly rather than described only on paper.
Published under the Shrewd Security organization in standard Transformers form, with an evaluation story and deployment path behind it.
Hugging FaceQuantized build
Llama-Phishsense-1B (GGUF)
A quantized build of the same model, packaged so it can be run locally without a hosted inference provider.
Deployment-oriented packaging: the point is that a defender can actually run the detector on their own hardware.
Hugging FaceThe papers are one output of the lab. The ledger tracks what happened after them.