Violet Teaming in Practice
A CSET pilot workshop tests violet teaming, a method integrating technical, security and policy views, to govern biodefense-relevant AI capabilities, surfacing non-obvious misuse pathways and implementation constraints.
Research / Sources
CSET
Research on AI governance and emerging technology. cset.georgetown.edu
Focus All
Topic Evals & standards
Date Any time
The Mechanisms of AI Harm: Lessons Learned from AI Incidents
Analyzes AI Incident Database reports to identify six mechanisms through which AI deployment causes harm, arguing mitigation needs diverse sociotechnical strategies beyond model-based approaches.
Harmonizing AI Guidance: Distilling Voluntary Standards and Best Practices into a Unified Framework
CSET distills over 7,000 recommended practices from 52 AI standards and guidance reports into a single 258-practice framework for organizations adopting AI, integrating safety and security.
How to Assess the Likelihood of Malicious Use of Advanced AI Systems
A CSET report proposes a framework for policymakers to assess the likelihood of malicious use of advanced AI, covering plausibility, performance, and observed-use evidence.
AI Incidents: Key Components for a Mandatory Reporting Regime
Proposes standardized key components for a mandatory AI incident reporting regime, covering incident type, harm severity, technical data, affected parties and context, drawing on transportation, healthcare and cybersecurity reporting.
Enabling Principles for AI Governance
CSET report arguing that effective AI governance requires tracking AI risks and incidents, building regulator AI literacy, and keeping policies adaptable as AI evolves.
An Argument for Hybrid AI Incident Reporting
Proposes a federated hybrid framework of mandatory, voluntary and citizen reporting to standardize and aggregate AI incident data, drawing lessons from other sectors' incident databases.
Comment on NIST RFI Related to the Executive Order Concerning Artificial Intelligence (88 FR 88368)
CSET's comment to NIST compiles recommendations on AI red-teaming, incident criteria, risk management and standards to help implement the AI Executive Order's safety requirements.
Repurposing the Wheel: Lessons for AI Standards
A CSET and CNAS workshop report drawing lessons from finance, worker safety, cybersecurity and medical-device standards to inform AI standards development and governance.
AI Incident Collection: An Observational Study of the Great AI Experiment
CSET explainer defining criteria for AI incident collection and comparing mandatory, voluntary and citizen reporting models for tracking real-world AI harms.
Adding Structure to AI Harm
CSET introduces a standardized conceptual framework for defining, classifying and tracking real-world harms caused by deployed AI systems, to support risk mitigation analyses.
One Size Does Not Fit All
CSET's research agenda argues AI assessment must cover diverse products and tools, and proposes studying testing, standards, auditing and certification to reduce AI safety and trust risks.
Comment to NIST on the AI Risk Management Framework
CSET's comment on NIST's second draft of the AI Risk Management Framework praises revisions to audience, risk definition and the Govern function, and flags interactions among AI systems as a future risk.
Classifying AI Systems
CSET brief developing and testing frameworks to classify AI systems along policy-relevant dimensions, comparing over 1,800 classifications to support risk assessment, governance and system management.
AI Accidents: An Emerging Threat
CSET policy brief classifying AI accident types (robustness, specification, assurance failures) with hypothetical scenarios and recommending policy actions to reduce accident risks from deployed ML systems.
AI Verification
Proposes concepts and technical mechanisms for verifying states' compliance with international arms control rules on military AI, including inspections and continuous verification.