AI Cyber Defense Act: the U.S. Prepares Federal Standards Against AI Model Theft | inicIA
Direct answer
On October 22, 2026, the AI Cyber Defense Act was introduced in the U.S. House of Representatives, a bipartisan bill from Reps. Pat Harrigan (R-N.C.) and George Whitesides (D-Calif.), reported exclusively by CyberScoop. The proposal tasks the NIST (National Institute of Standards and Technology) with developing voluntary industry standards to detect and prevent "distillation attacks" —training a model by copying the outputs of a more advanced one— and creates an AI Cybersecurity Roundtable for public-private collaboration. NIST would have 180 days after enactment to begin stakeholder engagement, and entities tied to China, Russia, Iran, North Korea and Cuba would be excluded from developing the standards.
What a "Distillation Attack" Is (and Why It Matters to Your Business)
- The technique: training a new model by massively querying a frontier model's outputs, replicating its capabilities at a fraction of the original development cost.
- The scale: U.S. officials and companies like Anthropic, OpenAI and Google have identified systematic campaigns where actors use burner accounts and false identities to extract data from frontier models while evading geographic restrictions.
- The congressman's line: Harrigan told CyberScoop that maintaining the U.S. industry's 6-to-12-month lead is "incredibly expensive," and adversaries find it more profitable to "stay 6 to 12 months behind the U.S., let the engine burn hot and, for 5 to 10% of the spend, distill and steal."
- The important nuance: NIST must first define what a distillation attack is and separate it from legitimate collaboration between open-weight and open-source models, which also drives innovation.
What to Prepare in Your Company (Even Though It's Still a Bill)
- Access controls on your AI APIs: if you expose models or agents to third parties, log who uses them, at what volume and with what patterns. Detecting anomalous prompt activity is the first line of defense.
- Extraction monitoring: spikes of systematic queries, mass "exam-style" question patterns or traffic from freshly created accounts are classic distillation-attempt signals.
- Diligence on your AI vendors: ask how they protect their models against extraction and which frameworks they follow. When NIST publishes standards, vendors already compliant will have the edge.
- Separate collaboration from exposure: if you contribute to open-source projects, define which weights, data and evaluations you share and which you don't. The bill aims to protect innovation without killing open collaboration.
The Strategic Read
- the U.S. chooses voluntary standards, not fines: like October's safety pact with big tech, Washington bets on guided self-regulation before hard regulation. For your company, that means standards become a competitive advantage before they become an obligation.
- Model cybersecurity is a boardroom topic: it's no longer just "don't leak customer data" — it now includes protecting your own models and agents as intellectual-property assets.
Related links
- Washington wants mandatory AI rules: the Cantwell plan and Trump's task force
- Trump signed an AI safety pact with no fines or teeth: what changes for your company
- The EU already fines for AI: up to €35M — what to know if you sell to Europe
Frequently asked questions
What does the AI Cyber Defense Act propose?
Tasking NIST with creating voluntary standards against "distillation attacks" (copying models by training on their outputs) and creating an AI Cybersecurity Roundtable to coordinate government with industry.
Who introduced it and when?
Reps. Pat Harrigan (R-NC) and George Whitesides (D-CA), on Oct 22, 2026. It's bipartisan and still a bill, not an enacted law.
What is a "distillation attack"?
The practice of training an AI model by massively querying a more advanced model's outputs, replicating its capabilities at a fraction of the cost. The bill targets actors from Russia, China, Iran and North Korea doing it.
What should my company do now?
Implement access controls and anomalous-usage monitoring on your AI APIs and agents, require anti-extraction measures from your vendors, and track the NIST standards publication to adopt them early.
