AI Technology Observations | 21 October 2026, 07:00

This observation covers technical facts published and verified on 19–20 October 2026. Two items are included: a model deployment restriction affecting government systems, and evidence of agent misuse in real-world network environments.

United States publishes federal civilian exclusion and removal order for DeepSeek

The US Federal Register formally published a Department of Homeland Security notice of order on 20 October. Issued under the Federal Acquisition Supply Chain Security Act, the order does two things. It excludes information technology products and services developed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd and affiliates as defined by the applicable rules from federal civilian agency procurement. It also requires civilian agency information technology to remove products from DeepSeek and its affiliates, as well as third-party products and services developed using DeepSeek implementations. The document entered public inspection on 19 October and was published as 91 FR 66346 on 20 October, so the new fact in this review is a public, verifiable federal document rather than a repetition of earlier restrictions imposed by individual agencies or states.

The order directly changes the conditions under which DeepSeek models and related third-party implementations may enter US federal civilian information systems and procurement. Reuters independently confirmed the procurement exclusion and removal scope. It also recorded DHS's stated risk grounds: possible transfer of sensitive data to China, production of actionable intelligence about US people or agencies, and possible malware or other security vulnerabilities in the models. The public material does not disclose all evidence supporting those assessments or a single public deadline for completing removal. The stated risks are therefore government findings, not model properties individually demonstrated through public technical testing. The order is not a blanket ban covering private users, every US business or worldwide use.

Technical investigation of AI agents used against live websites receives partial independent corroboration

NorForge published a technical investigation on 19 October based on publicly exposed Cairn control panels. Researchers identified more than 20 panels that were likely connected to malicious activity. The Chinese-language AIA0001 cluster comprised 14 panels and more than 1,100 attack projects, mainly targeting organisations in the United States, Japan, Australia and Canada. Panel records showed operators supplying a target and short objective, after which agents performed reconnaissance, vulnerability discovery, exploitation attempts and persistence actions. NorForge estimated an overall success rate of about 10 per cent, while AIA0001 projects that reached their goals took slightly less than 23 hours on average. These figures describe the exposed panels observed by the researchers and are not population-wide statistics for all AI-assisted attacks.

Compared with earlier reports of AI-assisted cyberattacks, the new evidence includes observable multi-agent execution records, specific exploit chains and validation against live systems. NorForge confirmed that an uploader web shell attributed to an agent remained accessible on a large hardware-parts supplier and validated an intrusion chain against a large US retailer. Reuters separately reviewed the activity, reported more than 1,200 attempted intrusions and more than 100 compromised websites, and independently identified targets including Deere, SharkNinja and Cummins. Deere said remediation had been completed and there was no indication that sensitive information was accessed. SharkNinja said the affected system was a standalone support site and that the incident had been contained, while Cummins was still investigating a potential attack on a third-party vendor's site. Worker names led NorForge to infer that some tasks used DeepSeek models or a Claude Code-style harness, but the model and invocation path were not independently confirmed for every task. The full extent of data extraction also remains unresolved.

Sources

US Federal Register, Notice of Order Regarding Exclusion and Removal of DeepSeek Products From Federal Civilian Executive Branch Information Technology Systems and Procurements, 20 October 2026, https://www.federalregister.gov/documents/2026/10/20/2026-21447/notice-of-order-regarding-exclusion-and-removal-of-deepseek-products-from-federal-civilian-executive

Reuters, US banning Chinese AI DeepSeek from government agency IT systems, 19 October 2026, https://www.reuters.com/technology/us-banning-chinese-ai-deepseek-government-agency-it-systems-2026-10-19/

NorForge, Inside Cairn: Open Windows into massive AI-Driven Cyber Operations, 19 October 2026, https://norforge.com/blog/inside_cairn

Reuters, How one hacker used AI agents and a few thousand dollars to break into scores of sites, 19 October 2026, https://www.reuters.com/legal/government/how-one-hacker-used-ai-agents-few-thousand-dollars-break-into-scores-sites-2026-10-19/