Brief · Both
Your AI options for incident response just got narrower
Hosted frontier models already refuse to help mid-breach. The usual fallback is an open-weight model you run yourself. For federal civilian agencies, one of the popular ones just got thrown out.
On Oct 19 at 8:45 AM ET, DHS filed a Federal Register notice, publishing Oct 20, for FASCSA Exclusion and Removal Order #2026-01. The order's original release date is July 24, 2026. It directs federal civilian agencies to "(1) exclude DeepSeek, DeepSeek-affiliated entities, and their products from civilian agency procurements, and (2) remove products provided by DeepSeek and DeepSeek entities, and third-party products and services developed using DeepSeek implementations, from civilian agency information technology systems."
What it covers, and what it doesn't:
- Who: federal civilian executive-branch agencies, except their national security and SCI systems. DoD and ODNI have their own authority.
- Contractors: under FAR 52.204-30, DHS FASCSA orders apply to non-DoD contracts. Contractors can't provide or use prohibited covered articles in performance unless a waiver applies.
- Not private companies in general. The order doesn't reach them.
- Not "all Chinese models." CyberScoop notes DeepSeek is the only Chinese model named.
- Not an approved list. It excludes DeepSeek-family products. It doesn't bless anything else.
- Exceptions: agencies had to send a written request within 60 days of issuance (the order's original release date was July 24, 2026, so the deadline was around Sept 22), with compelling justification. That window is closed. There's no built-in carve-out for research or security testing.
Until removal is finished, agencies "are encouraged" to monitor traffic in and out, enhance logging and exfiltration controls, and flag anomalous activity. Encouraged, not ordered.
My take: defenders are getting squeezed from both sides. Hosted models refuse the work, thanks to safety rules masquerading as security. Now the go-to open-weight fallback is off the table for federal civilian systems. This order doesn't bind anyone else, but federal rules have a way of showing up in contracts, vendor questionnaires and customer policies. I'd plan as if that's the direction, not wait to be told.
The incident-response lesson is the same one, with a deadline attached. Your IR model stack is a dependency. Pick it, approve it and test it before the breach. Finding out mid-incident that the model you've been relying on is a covered article for your customer is a bad way to spend hour three.
What to do:
- Inventory where AI models run in your IR workflow, hosted and local, including tools built on top of someone else's model.
- Federal civilian work, or a DHS-clause contract? Check for DeepSeek and third-party products built on DeepSeek implementations.
- Choose a self-hosted fallback that's approved for your environment, not just one that's good at not refusing. Get it signed off now.
- Check your customers' rules, not just yours. A model that's fine for you may be banned for the agency or contractor you're responding for.
- Log every refusal from hosted models, with model, timestamp and request ID, so you can make the case for defender access later.
- Agencies with DeepSeek still in place: monitor traffic, tighten logging and exfiltration controls, and flag anomalies until it's gone.
Sources
- Federal Register public inspection: DHS notice, FR Doc 2026-21447 (filed Oct 19, 2026)
- FAR 52.204-30
- CyberScoop: DHS bans DeepSeek from federal agencies
- Bloomberg Law: US moves to bar DeepSeek from federal agencies and contracts
- Reuters via Moneycontrol: US orders removal of DeepSeek from civilian government systems
- Quartz: DeepSeek ban for civilian agencies