Latest / AI Newspaper / DHS Orders DeepSeek Out of Federal Agencies, Including Tools Built on Its Models

Transcript

DHS Orders DeepSeek Out of Federal Agencies, Including Tools Built on Its Models · AI Newspaper

• 0:00
  Welcome to the AI Newspaper, your daily update
• 0:02
  on everything happening with AI. Visit AINewspaper
• 0:05
  .org to see today's stories and sign up for email
• 0:08
  alerts. Glad to be here for today's discussion.
• 0:10
  So today is October 20th, 2026, and we are looking
• 0:14
  at a stack of official documents that were published
• 0:17
  this morning. And honestly, they signal a monumental
• 0:20
  shift in how the United States government handles
• 0:22
  artificial intelligence. It really is a huge
• 0:25
  shift. We are looking at the official Federal
• 0:28
  Register publication of a sweeping new order,
• 0:30
  plus some urgent federal memos. Right. And it
• 0:33
  all points to one very stark reality. DeepSeek
• 0:36
  products are now completely banned from federal
• 0:38
  civilian executive branch agencies. The magnitude
• 0:41
  of this publication really cannot be overstated.
• 0:43
  I mean, by ordering DeepSeek out of federal civilian
• 0:46
  systems, Washington is doing something entirely
• 0:49
  new. They are going after other companies' tools,
• 0:53
  too, right? Exactly. They're ordering it out
• 0:56
  of tools built on its models. So Washington is
• 0:59
  essentially turning the geographic origin of
• 1:01
  an AI model into a strict supply chain question.
• 1:04
  Which means every vendor selling software to
• 1:06
  the government now has to answer for the foundational
• 1:08
  AI operating inside their products. Right. It
• 1:11
  completely rewires the government's relationship
• 1:14
  with software vendors. So we are going to break
• 1:16
  down the mechanics of this ban, the underlying
• 1:18
  cybersecurity threats driving it, and, well,
• 1:21
  the immediate scramble it's causing for IT departments
• 1:23
  across the country. Because when you look at
• 1:25
  the text of the order, there is a hidden clause
• 1:27
  here that is going to cause absolute chaos for
• 1:30
  the software industry. Yeah, it goes far beyond
• 1:32
  an IT administrator just deleting a chatbot app
• 1:35
  from a government -issued smartphone. To understand
• 1:38
  the true scope, we really have to look at the
• 1:40
  specific mechanism of the document published
• 1:42
  today. Which is the Federal Acquisition Supply
• 1:44
  Chain Security Act Exclusion and Removal Order,
• 1:47
  right? Yes, specifically order number 2026 -01.
• 1:51
  And it's interesting because it was actually
• 1:53
  signed months ago. Oh, really? When was that?
• 1:55
  On July 24th, 2026 by Homeland Security Secretary
• 2:00
  Mark Wayne Mullen. It was filed publicly yesterday.
• 2:03
  And as of today, it officially hits the books.
• 2:06
  So it's legally binding right now. Exactly. The
• 2:09
  immediate mandate is that all federal civilian
• 2:11
  executive branch agencies must immediately exclude
• 2:14
  Hangzhou Deep Seek Artificial Intelligence Company
• 2:16
  from all procurements. And its affiliates, I
• 2:19
  assume. Yep. Affiliates, too. Wow. So that includes
• 2:22
  agency -approved product lists, service provider
• 2:25
  equipment lists, e -commerce markets, just everything.
• 2:29
  Everything. But the detail in this Federal Register
• 2:31
  publication that completely alters the landscape
• 2:34
  is the third -party provision. Ah, right. The
• 2:38
  third -party rule. This is the big one. It is.
• 2:40
  The ban explicitly includes third -party products
• 2:43
  and services that are developed using DeepSeek
• 2:45
  implementations. So meaning if a vendor sells
• 2:48
  a seemingly unrelated piece of software to a
• 2:51
  government agency, they are on the hook. Exactly.
• 2:54
  Maybe it's a routine data entry tool or a logistics
• 2:57
  tracker. Or like a document summarization program.
• 3:01
  But if it runs on a deep -seek model anywhere
• 3:03
  under the hood, that software is now strictly
• 3:05
  prohibited. Yep. The entire burden of proof just
• 3:08
  shifted to the vendor. You can't just slap your
• 3:10
  company's logo on a software dashboard and sell
• 3:12
  it to an agency anymore. No, you really can't.
• 3:14
  The government is forcing vendors to audit their
• 3:16
  own dependencies now. So if you integrated DeepSeek's
• 3:19
  latest update, say the v4 .1 -Flash model that
• 3:23
  shipped recently, because it's highly capable
• 3:25
  and cuts your server costs, what happens? Your
• 3:27
  entire application is legally barred from civilian
• 3:31
  agency IT systems. Full stop. I want to make
• 3:34
  sure everyone listening visualizes the reality
• 3:36
  of this supply chain problem. It's a tricky concept
• 3:40
  to grasp at first. Right. So it is like a massive
• 3:44
  office building manager issuing a total ban on
• 3:47
  a very specific brand of electronic security
• 3:49
  lock. Oh, that's a good analogy. Yeah. So the
• 3:52
  ban covers the lock sold directly by that brand,
• 3:54
  obviously. But the new rule also covers any door
• 3:57
  sold by any other company if that door happens
• 4:01
  to have that specific lock mechanism embedded
• 4:03
  inside it. Exactly. You can't tell the building
• 4:07
  manager, well, hey, I bought this door from a
• 4:09
  totally different American company. Right, because
• 4:11
  if the prohibited lock is buried inside the wood,
• 4:13
  the entire door has to be removed. That perfectly
• 4:16
  illustrates the dilemma, especially because in
• 4:18
  modern software development, AI models are incredibly
• 4:21
  easy to hide inside those doors. Developers just
• 4:25
  plug AI models into their applications via APIs
• 4:27
  all the time. All the time. But the origin of
• 4:30
  that foundational model is no longer just a technical
• 4:33
  feature. No, it is a critical component of the
• 4:35
  national security supply chain now. Let's clarify
• 4:38
  the concept of an API for any listeners who might
• 4:40
  not be software developers. Good idea. An API
• 4:43
  or application programming interface is essentially
• 4:46
  a digital courier surface. Right. It's just passing
• 4:49
  messages back and forth. So if I build an app
• 4:51
  that summarizes long PDFs, my app doesn't actually
• 4:54
  do the heavy lifting. When a user uploads a PDF,
• 4:57
  my app uses an API to silently send that document
• 5:00
  out through the Internet to an AI model's servers.
• 5:03
  The AI reads it, generates the summary, and then
• 5:06
  the API carries the summary back to your app
• 5:08
  to show the user. Exactly. So if a vendor is
• 5:11
  using a DeepSeek API, they are establishing a
• 5:14
  direct pipeline between the federal government's
• 5:16
  data and DeepSeek's servers. And that digital
• 5:19
  pipeline is precisely where the Department of
• 5:22
  Homeland Security begins its threat assessment.
• 5:24
  Because of where those servers are located, right?
• 5:26
  Yeah. The order outlines several specific vectors
• 5:30
  and they are heavily focused on data transit
• 5:32
  and the mechanics of generative AI. Right. DHS
• 5:36
  points out that DeepSeek servers are located
• 5:38
  in the People's Republic of China. And under
• 5:41
  the PRC's legal system, government cyber and
• 5:44
  intelligence functions are granted extensive
• 5:46
  authority. Authority to access private company
• 5:49
  infrastructure, basically. Yeah. And any data
• 5:52
  residing on it. Which means the fundamental prompt
• 5:54
  and response process of a chatbot becomes a massive
• 5:57
  intelligence vulnerability. You are literally
• 5:59
  handing the model the exact data you want it
• 6:01
  to process. Right. If a U .S. government employee
• 6:04
  or even a contractor uses a third party tool
• 6:06
  to summarize internal meeting notes or write
• 6:09
  software code, that data is transiting out of
• 6:12
  the agency. It's transiting into a jurisdiction
• 6:13
  where the U .S. government has zero visibility
• 6:16
  or control. And the order explicitly warns about
• 6:19
  the aggregation of this data, too. Yeah, that's
• 6:21
  a key point. Even if an individual user thinks
• 6:24
  they are just asking a generic question or summarizing
• 6:27
  a mundane, unclassified report. The sheer volume
• 6:30
  of queries coming from a federal agency over
• 6:32
  time paints a very detailed picture. Right. Combined
• 6:36
  with publicly available information, these queries
• 6:39
  could produce actionable intelligence on U .S.
• 6:42
  citizens. Or government priorities or internal
• 6:44
  programs. Exactly. Additionally, for the versions
• 6:48
  of the models that DeepSeq hosts directly, the
• 6:51
  order notes the account creation process collects
• 6:53
  personally identifiable information. And they
• 6:56
  also mention it may engage in digital fingerprinting.
• 6:58
  Yes, which is a major red flag for DHS. Let's
• 7:02
  define digital fingerprinting for a second because
• 7:04
  it sounds pretty alarming. It is. It means the
• 7:07
  software isn't just answering your prompt. It's
• 7:10
  actively gathering subtle clues about your device,
• 7:12
  like your operating system, your IP address,
• 7:15
  your browser settings. Even your screen resolution.
• 7:17
  It uses all that to create a unique, trackable
• 7:20
  profile of you, the user. Even if you are trying
• 7:23
  to remain anonymous. That is the exact concern.
• 7:26
  Once a federal employee or contractor is fingerprinted,
• 7:29
  their activity can be tracked. correlated across
• 7:32
  multiple sessions or platforms, creating this
• 7:34
  escalating security risk for the agency they
• 7:37
  work for. Right. I do have to push back on this
• 7:39
  logic, though. Okay, let's hear it. Let's say
• 7:41
  I'm a savvy software vendor. I know all about
• 7:44
  the data transit risks and the API pipelines
• 7:46
  sending information to China. So I just don't
• 7:49
  use the API. Okay. Instead, I download the model's
• 7:53
  open weights, meaning I legally download the...
• 7:56
  underlying mathematical blueprint of the AI brain,
• 7:59
  and I install it entirely on my own highly secure
• 8:02
  US -based servers. So you're running it locally.
• 8:05
  Exactly. It is completely disconnected from the
• 8:07
  outside internet. There is zero data transiting
• 8:11
  back to the PRC. The pipeline is completely severed.
• 8:13
  Right. Why is DHS still treating my offline,
• 8:16
  isolated application like a loaded weapon? That
• 8:20
  is a critical distinction, and DHS actually addresses
• 8:22
  it directly in their technical assessment. What
• 8:24
  do they say? The risk is not purely geographical
• 8:27
  data transit. The risk also lies in the code
• 8:30
  itself. Wait, the code itself is the problem?
• 8:34
  Yes. The DHS order states that the models themselves,
• 8:37
  even when downloaded and run locally, may contain
• 8:40
  malware or severe security vulnerabilities. Wow.
• 8:43
  And to support this, they cite extensive open
• 8:46
  source security research. Ah, the now secure
• 8:48
  research from back in February 2025. Let's unpack
• 8:51
  what they actually found because it kind of...
• 8:53
  of undercuts the idea that running the model
• 8:55
  on your own hardware makes you safe. It completely
• 8:57
  undercuts it. Now Secure conducted an analysis
• 9:00
  of the DeepSeek iOS mobile app. And what did
• 9:03
  they find? They found multiple critical security
• 9:06
  and privacy flaws embedded in the software's
• 9:08
  architecture. So when you take a model with documented
• 9:11
  fundamental code vulnerabilities and you embed
• 9:14
  it into your proprietary software, you're in
• 9:17
  trouble. You are importing those vulnerabilities
• 9:19
  straight into your own ecosystem. The vulnerability
• 9:22
  inherits the access level of the application
• 9:24
  it's embedded within. Precisely. So if I build
• 9:28
  a secure data management tool for, say, the Department
• 9:31
  of Energy, and I give my tool deep administrative
• 9:33
  access to the network so it can do its job, but
• 9:36
  I've hidden a flawed, vulnerable, open -weight
• 9:40
  AI model inside it, that flaw now has administrative
• 9:44
  access to the Department of Energy. And that
• 9:46
  inheritance of risk is exactly why the third
• 9:49
  -party ban is so absolute. DHS just cannot allow
• 9:53
  deeply embedded compromised code to sit inside
• 9:55
  vendor applications. Regardless of whether those
• 9:58
  applications are phoning home to China or operating
• 10:00
  locally in a basement. And if you're an IT director
• 10:03
  listening to this right now, the logistical reality
• 10:06
  of this order is just terrifying. It really is.
• 10:10
  We are talking about tearing deeply embedded
• 10:12
  infrastructure out of enterprise systems basically
• 10:15
  overnight. You don't just hit uninstall on a
• 10:18
  foundational AI model that powers hundreds of
• 10:21
  daily automated workflows across an entire federal
• 10:24
  agency. No, it's woven into everything. So how
• 10:27
  are agencies actually supposed to execute this
• 10:29
  extraction without completely breaking their
• 10:31
  own operations? Well, the authors of the order
• 10:34
  clearly understand the logistical nightmare this
• 10:36
  creates, which is why they built in a highly
• 10:38
  structured pressure valve. OK, what kind of pressure
• 10:41
  valve? While the ban is effective immediately
• 10:43
  upon today's publication, executive agencies
• 10:46
  have a 60 -day window to formally request exceptions
• 10:49
  in writing. Oh, okay. And those requests are
• 10:52
  not reviewed internally, right? They go straight
• 10:54
  to CISA? Yes, the Cybersecurity and Infrastructure
• 10:58
  Security Agency, specifically to their Federal
• 11:00
  Acquisition Supply Chain Security Act Order Implementation
• 11:04
  Team. That's a mouthful. But the threshold for
• 11:07
  getting an exception approved is incredibly high,
• 11:09
  isn't it? Oh, incredibly high. An agency cannot
• 11:12
  simply complain that removing the software is
• 11:14
  expensive or annoying. Right. They must provide
• 11:16
  a compelling, documented justification proving
• 11:19
  how immediate removal would severely impact mission
• 11:22
  critical functions. Or they must cite major national
• 11:26
  security considerations. And even if they can
• 11:29
  prove that, they are not allowed to simply maintain
• 11:31
  the status quo. Not at all. They have to outline
• 11:33
  alternative mitigations. They have to tell CISA
• 11:36
  exactly what defensive measures they are putting
• 11:38
  in place to neutralize the risk of keeping the
• 11:40
  prohibited software running. Right. Think of
• 11:43
  it like discovering a crumbling, compromised
• 11:45
  concrete foundation underneath a massive office
• 11:48
  building. But you can't evacuate the building
• 11:50
  because the employees are doing mission -critical
• 11:52
  work. Exactly. You can't just rip the foundation
• 11:55
  out or the building collapses. So you have to
• 11:57
  install temporary support beams, put up heavy
• 11:59
  caution tape, set up 24 -7 security cameras,
• 12:02
  and build containment tarps. Just to keep the
• 12:05
  damage from spreading while you slowly figure
• 12:06
  out how to extract the bad concrete safely. And
• 12:09
  the interim cybersecurity measures mandated by
• 12:12
  DHS look exactly like that heavy caution tape
• 12:14
  and temporary support structure. They do. During
• 12:17
  this transitional period before full removal
• 12:19
  is achieved, agencies are instructed to implement
• 12:22
  severe restrictions. Like monitoring the flow
• 12:25
  of network traffic in and out of the compromised
• 12:27
  systems continuously. And they have to implement
• 12:30
  enhanced logging. Meaning every single digital
• 12:33
  action taken by the software is recorded and
• 12:35
  heavily scrutinized. Every single one. They also
• 12:38
  have to add strict controls to prevent data exfiltration.
• 12:41
  Right, exfiltration being the technical term
• 12:43
  for sneaking data out the back door. If the AI
• 12:46
  model tries to quietly copy a batch of files
• 12:48
  and send them to an unauthorized external server,
• 12:50
  the systems must be configured to instantly block
• 12:53
  that transfer and trigger an alarm. And CISA
• 12:56
  knows that not every civilian agency has the
• 12:58
  internal cyber capabilities to set up that level
• 13:01
  of tripwires overnight. Which is why CISA has
• 13:03
  offered to deploy technical assistance teams
• 13:06
  to help agencies comply with these strict monitoring
• 13:08
  requirements. The oversight is just designed
• 13:11
  to be relentless here. Truly. The Secretary of
• 13:14
  Homeland Security is actually legally required
• 13:16
  to review this entire order within one year.
• 13:18
  And conduct annual reviews after that to ensure
• 13:21
  compliance isn't slipping. Yeah. This 60 -day
• 13:24
  scramble for civilian IT directors is intense,
• 13:27
  but to really grasp the structural forces shaping
• 13:30
  the tech industry, we have to zoom out a bit.
• 13:32
  We do. This order did not just drop out of a
• 13:35
  clear blue sky. No. We are watching a patchwork
• 13:38
  of localized restrictions rapidly stitch itself
• 13:41
  into a massive, unavoidable blanket. Right. And
• 13:44
  the jurisdiction of this specific Federal Register
• 13:47
  publication is important to clarify. It applies
• 13:49
  exclusively to federal civilian executive branch
• 13:52
  agencies. Exactly. It explicitly exempts the
• 13:56
  Department of Defense, which, interestingly,
• 13:59
  the text refers to using its historical name,
• 14:01
  the Department of War. Oh, that is interesting.
• 14:03
  And it also exempts the intelligence community,
• 14:06
  national security systems, and secure compartmented
• 14:09
  information systems. But they aren't exempted
• 14:11
  because they are allowed to use the software.
• 14:13
  Right. They are exempted because they have their
• 14:16
  own much faster authorities and they already
• 14:18
  banned it. Decisively. Congress stepped in and
• 14:21
  restricted the use of DeepSeek in intelligence
• 14:23
  community systems back in December 2025. And
• 14:26
  the Defense Information Systems Agency, DISA,
• 14:28
  blocked access to the models across military
• 14:30
  networks around that same time. We also saw action
• 14:33
  on the civilian device side early in 2025. When
• 14:35
  the Commerce Department and the Navy banned the
• 14:38
  app on government -issued smartphones. And at
• 14:40
  the state level, governments in New York, Virginia,
• 14:42
  and Tennessee implemented their own bans on government
• 14:45
  devices. So the evolution from those earlier
• 14:48
  actions to today's order is really the crux of
• 14:50
  the issue for the software market. It is. Banning
• 14:54
  an app on a government smartphone is a device
• 14:56
  -level restriction. It is relatively easy to
• 14:59
  enforce through a mobile device management platform.
• 15:01
  You just block the app download. But what this
• 15:03
  new FASCA order does is scale that philosophy
• 15:07
  up to the supply chain level. Exactly. It blocks
• 15:10
  the foundational technology from entering the
• 15:12
  entire civilian purchasing ecosystem, no matter
• 15:15
  how deeply it is buried inside other tools. That
• 15:17
  is the real paradigm shift. Washington is actively
• 15:20
  turning where an AI model comes from into a mandatory
• 15:23
  supply chain audit. If you are a vendor selling
• 15:27
  software to the government, the ingredients of
• 15:29
  your code now matter just as much as what the
• 15:31
  code actually does. You can have a tool that
• 15:33
  performs a perfectly benign, highly efficient
• 15:35
  function. But if the AI engine processing the
• 15:38
  data is sourced from a prohibited entity, the
• 15:41
  utility of your tool is just irrelevant. Your
• 15:43
  product is banned. And this is part of a much
• 15:46
  wider, aggressive net being cast by the federal
• 15:48
  government over technology originating from China.
• 15:51
  The AI supply chain is really just the latest
• 15:53
  front. For example, the Federal Communications
• 15:56
  Commission is reportedly preparing to vote on
• 15:58
  barring Chinese laboratories from testing electronics
• 16:01
  destined for the U .S. market. That is a massive
• 16:04
  bottleneck considering a huge percentage of U
• 16:07
  .S. electronics currently undergo compliance
• 16:09
  testing in those exact labs. We also recently
• 16:11
  saw the FCC ban the import of new models of Chinese
• 16:15
  drones, as well as consumer Wi -Fi routers. The
• 16:19
  perimeter of what is considered acceptable technology
• 16:21
  infrastructure is shrinking rapidly. Which is
• 16:25
  why the developers, executives, and IT professionals
• 16:28
  listening to this need to understand that the
• 16:30
  era of casually grabbing the cheapest or fastest
• 16:34
  open -weight AI model off the internet is over.
• 16:37
  You can't just plug it into enterprise software
• 16:40
  anymore. You have to know the geopolitical footprint
• 16:42
  of your code. You have to audit every single
• 16:45
  software dependency in your stack. Because if
• 16:48
  a federal agency is forced by DHS to audit their
• 16:51
  systems and they discover that your proprietary
• 16:53
  software is secretly making an API call to a
• 16:57
  prohibited model. Or running a vulnerable open
• 17:00
  weight model locally. Right. You are not just
• 17:02
  going to lose a software contract. You are going
• 17:05
  to be caught up in a federal supply chain security
• 17:07
  violation. You basically become the threat vector.
• 17:09
  The compliance burden this places on the private
• 17:11
  sector is just immense. Vendors will have to
• 17:14
  maintain meticulous records, essentially creating
• 17:17
  an AI bill of materials for every piece of software
• 17:20
  they sell. Proving the origin and security of
• 17:22
  every model they utilize. Exactly. This leaves
• 17:25
  us with a massive question about the future of
• 17:27
  the technology market for you to think about.
• 17:30
  If Washington is now treating the origin of an
• 17:32
  AI model as a strict, heavily audited supply
• 17:35
  chain requirement, how long will it take for
• 17:37
  the broader private sector to follow suit? Think
• 17:40
  about major banks. sprawling hospital networks
• 17:42
  or critical energy infrastructure providers.
• 17:45
  They face the exact same data exfiltration and
• 17:48
  malware risks that the government does. So how
• 17:51
  long until a private hospital network demands
• 17:53
  this exact same AI bill of materials from their
• 17:56
  vendors before signing a contract? Are we watching
• 17:58
  the tech industry be forced to divide itself
• 18:00
  entirely into a heavily audited government -cleared
• 18:03
  software ecosystem and a less secure commercial
• 18:06
  -only ecosystem? If the liability of a third
• 18:10
  -party model compromising an entire enterprise
• 18:12
  network becomes too great, bifurcation of the
• 18:15
  market might be the only viable path forward
• 18:17
  for major software vendors. It is a profound
• 18:19
  structural change, fundamentally altering how
• 18:22
  software is built and sold. And the rules are
• 18:25
  being written in real time. Thanks for listening
• 18:27
  to the AI Newspaper. Visit AINewspaper .org to
• 18:31
  catch up on the latest updates and sign up for
• 18:32
  free email alerts.