SAIS Labs
Join the Wait List
SAIS Labs
Secure AI Systems Labs · Prime Federal Solutions

We research, build and test secure AI systems, with published results — creating trusted AI security tools with federal grade security and frontier level capability.

Secure [AI] for all humanity

What the lab does

AI security as a
systematic science.

AI attacks AI. We measure what breaks, fix it, prove the fix and share the method. Then we run it again.

Method published with every result
Including what did not work
Findings that hold up become controls in AXIS
01QuestionAI SECURITY +THREAT RESEARCH02AttackTESTING +EVALUATION03MeasureMODEL EVALUATIONENVIRONMENT04FixGUARDRAILS +SAFEGUARDS05ProveAI GOVERNANCE +COMPLIANCE06ShareAI READINESS +WORKFORCEAI VS AIRED AITARGETBREACHHELDWhat could break?Red AI attacks the targetScore what got throughThe finding becomes a controlRe-test. The control holds.Publish the method and resultROUND 01
RED AI vs TARGETROUND 01
01
Question
What could break?
AI Security + Threat Research
02
AttackBREACH
Red AI attacks the target
Testing + Evaluation
03
Measure
Score what got through
Model Evaluation Environment
04
Fix
The finding becomes a control
Guardrails + Safeguards
05
ProveHELD
Re-test. The control holds.
AI Governance + Compliance
06
Share
Publish the method and result
AI Readiness + Workforce
Our services

The Trusted AI Approach

Driven by OMB M-25-21, NIST AI control overlays and FedRAMP 20x.Spending to secure AI is set to reach $4.8B in 2027 (Gartner).
Click a service for more detail.
TRUSTED
AI
Federal

AI Governance + Compliance

Trust that your AI is on record and cleared to run
Build your AI use-case inventory and flag high-impact AI
Prepare AI authorization (ATO) packages mapped to NIST SP 800-53 AI overlays and FedRAMP 20x
Guide AI buying with evaluation criteria for contracting officers
Set deployment thresholds and go/no-go criteria
Translate risk into playbooks your teams can run

WHY NOW

Every agency must list its AI and apply risk steps to high-impact uses (OMB M-25-21).
Commercial

AI Governance + Compliance

Trust that your AI is on record and cleared to run
Stand up an AI governance program aligned to NIST AI RMF and ISO/IEC 42001
Prepare for EU AI Act high-risk obligations
Score the AI tools your vendors bring in
Set deployment thresholds and go/no-go criteria
Translate risk into playbooks your teams can run

WHY NOW

63% of companies studied have no AI governance policy (IBM, 2025). EU AI Act high-risk rules arrive in late 2027.
Federal

Testing + Evaluation

Trust that you know how your AI behaves before it goes live
Test high-impact AI before it goes live
Assess AI independently (IV&V) before launch and each year after
Red-team systems to find failures
Validate AI for specific federal use cases
Run repeatable evaluations

WHY NOW

High-impact AI must be tested before launch. NIST CAISI and GSA are building evaluation into federal buying.
Commercial

Testing + Evaluation

Trust that you know how your AI behaves before it goes live
Red-team agents across multi-step attacks and tool use
Test MCP servers and tool permissions
Score agents against industry and our custom security benchmarks and test campaigns
Measure real-world impact
Run repeatable evaluations

WHY NOW

Gartner expects over half of successful attacks on AI agents to use access control gaps and prompt injection by 2029.
Federal

Guardrails + Safeguards

Trust that your safeguards hold in production
Test safeguards in production
Monitor deployed systems
Protect PII, PHI and CUI in AI data
Find shadow AI across networks and cloud
Provide managed AI assurance with ongoing reporting

WHY NOW

Agencies must keep watching high-impact AI after it goes live.
Commercial

Guardrails + Safeguards

Trust that your safeguards hold in production
Deploy safeguards through the AXIS gateway
Monitor deployed systems
Protect sensitive data in prompts and outputs
Sweep for shadow AI
Provide managed AI assurance services

WHY NOW

Shadow AI added about $670K to the average breach (IBM, 2025). AI usage control is the fastest-growing AI security segment (Gartner).
Federal

AI Security + Threat Research

Trust that you see attacks before they reach you
Emulate real threat actors
Identify emerging AI threats
Review agent security against OWASP GenAI ASI
Trace AI supply chains and where models come from
Respond to AI incidents with forensics

WHY NOW

The AI Action Plan calls for secure-by-design AI and federal AI incident response.
Commercial

AI Security + Threat Research

Trust that you see attacks before they reach you
Publish an AI threat intel feed of real attacks
Name new attack types and disclose AI flaws (CVEs)
Emulate real threat actors
Respond to AI incidents on retainer, with forensics
Release open-source AI security test tools
Shape standards with OWASP, CoSAI and NIST

WHY NOW

AI attacks change faster than standard security tools can keep up. Our research tracks new attacks as they appear and turns them into defenses you can use.
Federal

AI Readiness + Workforce

Trust that your people are ready to lead with AI
Prepare leaders to make AI decisions
Decide where AI adds value
Set rules for responsible use
Lead teams through AI-driven change

WHY NOW

Agencies must build AI skills and choose where AI fits the mission.
Commercial

AI Readiness + Workforce

Trust that your people are ready to lead with AI
Brief executives and boards on AI risk
Decide where AI adds value
Set rules for responsible use

WHY NOW

Leaders get a clear view of where AI pays off and where the risk sits, so they can invest with confidence and move first.
Federal

Model Evaluation Environment

Trust that the right model is always in place
Evaluate agentic and multimodal systems
Benchmark models against a capability floor
Score backup models so a swap is ready when a provider fails

WHY NOW

Built on AXIS. Tests run only on models your agency authorizes.
Commercial

Model Evaluation Environment

Trust that the right model is always in place
Evaluate agentic and multimodal systems
Benchmark models against a capability floor
Score backup models for fast swaps

WHY NOW

Built on AXIS, so testing runs where the work runs.
Safety Over SpeedTransparency OverBlack BoxesTrust ThroughCollaboration, NotIsolationMission BeforeMachineFairness isNon-NegotiableContinuousOversight,ContinuousImprovementMeasurable ValueOver TheoreticalPotentialResilience OverFragilityContinuousImprovement01TrustworthinessAI must earn confidencethrough transparency,fairness, explainability,and measurable value.02SecurityAI must be protected atevery layer — data,models, infrastructure,and human workflows.03SafetyAI must behave reliably,predictably, andresponsibly in the realworld.04HumanAccountabilityHumans remain responsiblefor decisions, outcomes,and oversight.TRUSTEDAI
TRUSTED AI
01
Trustworthiness
AI must earn confidence through transparency, fairness, explainability, and measurable value.
02
Security
AI must be protected at every layer — data, models, infrastructure, and human workflows.
03
Safety
AI must behave reliably, predictably, and responsibly in the real world.
04
Human Accountability
Humans remain responsible for decisions, outcomes, and oversight.
Nine more commitments
Safety Over SpeedTransparency Over Black BoxesTrust Through Collaboration, Not IsolationMission Before MachineFairness is Non-NegotiableContinuous Oversight, Continuous ImprovementMeasurable Value Over Theoretical PotentialResilience Over FragilityContinuous Improvement
Principles

The Trusted AI Framework

Why an assessment layer has to exist

An AI company that regulates its own security carries a conflict. The same people are heavily incentivized on being the industry model provider leader and on how thoroughly they solve for its negative impacts and vulnerabilities.

A trusted AI system must be reliably secure first. It doesn’t matter how capable the model is, if the negative ramifications of using it are too great.
Four pillars define Trusted AI. Nine more commitments orbit them.
Read all thirteen principles →
The work

Active Testing, Contributions and Analysis

Standards contribution

Contributions to the OWASP GenAI security work, published under our own names and freely available.

Securing Agentic Applications
GenAI Red Teaming Manual
GenAI Red Teaming Guide
Coming soon
AIR GAP

The evaluation lab

An air-gapped environment for testing model and agent behavior that should not be tested on a network.

Real and checkable. Where a finding needs an isolated environment to be honest, it gets one.

Applied research

Two filings, described by what they do rather than by claim language.

AI supply-chain hot-swapping
Automatically applied OWASP ASI controls
Patent pending · Provisionals filed

Latest Results - Red-teaming AXIS with models

Frontier models are evaluated on how well they attack AXIS itself — prompt injection, goal hijack, tool abuse, and getting a workflow published that should have been refused.

A model that gets through is a finding against us which immediately generates a fix and patch for our platform team to review and implement.