Shrewd Research is the reference standard for AI security, the way the metre is the reference for measurement. A safety claim should not be a vendor's assertion; it should trace to tamper-evident evidence an independent party can verify, replicate, and revoke. We produce that evidence: which containment properties hold, which fail under attack, and which assumptions cannot survive a capable adversary.
A strong negative result is useful. It can stop a lab from scaling a defense that cannot hold, keep a benchmark from becoming a misleading standard, redirect funding toward viable approaches, and prevent policy from encoding a technically false premise.
Study the fundamental trade-offs and failure boundaries of proposed AI defenses. We ask what a control can guarantee, what assumptions it requires, and where those guarantees become impossible.
Deliverable: A rigorous result that rules out a bad direction, narrows the viable design space, or identifies the assumptions a defensible approach must satisfy.
Design experiments that pressure-test models, safeguards, evaluation harnesses, and AI infrastructure against adaptive attacks rather than static benchmarks.
Deliverable: Reproducible evidence showing where a system fails, how reliably it fails, and which conclusions the data does and does not support.
Translate an important safety or security claim into a falsifiable test, then evaluate it independently against realistic threats and deployment conditions.
Deliverable: A clear pass, fail, or conditional result with the evidence and assumptions needed for others to challenge or reproduce it.
Replicate influential findings, build harder evaluations, and test whether reported results survive new models, stronger adversaries, and changed environments.
Deliverable: An independent replication, open benchmark, or evaluation method that makes a disputed claim easier to verify.
Turn technical findings into precise implications for policy, standards, funding, evaluations, and infrastructure design without overstating what the evidence proves.
Deliverable: A decision memo, technical standard input, or evidence-backed recommendation that separates established facts from open questions.
Verify, Record, Investigate, Audit, and Attest are applications of the research. They turn abstract findings about what can fail into tests and evidence that operators and independent evaluators can use.
Attack the system and validate its security claims.
Automated adversarial testing of training and inference environments. Instead of questionnaires and architecture diagrams, we continuously attack the environment and produce evidence about which security claims actually hold.
Before deployment, after major infrastructure changes, or when a provider's isolation claims need independent validation.
Reproducible attack paths, pass/fail results for each claim, evidence, and prioritized remediation.
Preserve trustworthy evidence of what actually happened.
A tamper-evident flight recorder for AI systems. Every consequential event is bound to the model, run, identity, tools, safeguards and authorization context that produced it.
When existing logs cannot reliably connect model behavior to infrastructure state, identities, safeguards, and human decisions.
An evidence model, collection architecture, integrity controls, and implementation plan.
Reconstruct failures automatically.
The interesting part is not logging. It is reconstructing the incident: timeline, initiating event, violated boundary, blast radius, likely causal chain and confidence-ranked hypotheses — cryptographically distinguishing observed facts from inferred conclusions.
During or after suspected containment failure, credential compromise, cross-tenant activity, or unexplained model behavior.
A defensible timeline, blast-radius analysis, evidence-backed hypotheses, and recommended containment actions.
Let independent evaluators verify models securely.
Cryptographically verifiable clean rooms for frontier-model evaluations. A lab provides encrypted model artifacts, an evaluator provides confidential tests, and the evaluation runs in an attested environment neither side has to fully trust.
When a model developer and evaluator need trustworthy results without disclosing frontier weights or private evaluation methods.
A constrained evaluation environment plus signed evidence of the model, configuration, controls, and results.
Issue an assurance result people can actually trust.
Not a certificate someone earned twelve months ago. An assurance state backed by continuously produced technical evidence — and revoked automatically when a property it depends on changes.
When customers, partners, insurers, or regulators need current evidence that specific security properties still hold.
An evidence-backed assurance result tied to explicit claims, dependencies, and revocation conditions.
Every important fact about a model — provenance, training environment, hardware, checkpoints, evaluations, deployment configuration, permissions, safeguards and runtime events — becomes signed, tamper-evident evidence. All five products are different views over the same graph.