Government agencies are rapidly exploring generative AI to modernize FOIA responses, investigations, and eDiscovery workflows. However, as these capabilities expand, so do threats and security breaches to the cybersecurity ecosystem.
Now is the time for agencies to reassess preexisting assumptions about what security readiness and resiliency look like in sensitive FOIA and eDiscovery environments. As agencies vet vendor partners, consider the five questions below to keep security as the foundation of every workflow.
Unpacking the Evolving Threat Landscape
Recent debate over frontier AI models has raised emerging questions about how government agencies should prepare for increasingly sophisticated cyber threats. Prior to the rise of AI, it would take months for humans to discover vulnerabilities, develop exploits, and execute an attack. But today, anyone with access to advanced AI tools — both good and bad actors — can become a threat and accomplish that work in mere hours or days.
Research from the U.K.’s AI Security Institute (AISI) warned that Anthropic’s Claude Mythos could autonomously carry out advanced, multi-step cyberattacks that would previously require weeks of work by human experts. A report from CrowdStrike found that AI-enabled hackers increased attacks by 89% in 2025, with the technology dramatically accelerating the speed at which hackers work. Brute-force attacks, sophisticated phishing campaigns, and convincingly fake websites made with AI add to the risk.
For many years, FedRAMP® Moderate was considered sufficient to satisfy procurement or compliance requirements. That’s no longer the case. Now, autonomous AI agents can work 24/7 to pinpoint vulnerabilities, generate exploits, and accelerate attacks at unprecedented speeds. As a result, agencies need to ask whether their vendor’s security baseline adequately reflects today’s risky environment.
The stakes remain high. Without stringent security, government agencies are up against enormous risk, across blown FOIA exemptions, compromised investigations, and defensibility issues that could result in adverse citations in court.
5 Questions to Keep Security as the Foundation
1. Does the vendor’s security authorization match the sensitivity of the data?
Across sensitive FOIA, investigation, and eDiscovery workloads, FedRAMP High needs to be the procurement floor, given how quickly the threat landscape is evolving. These workflows contain sensitive records and privileged communications that could materially affect litigation and defensibility. Even if the data is classified as “moderate impact,” the risk is not worth it.
Understanding your vendor’s security safeguards is critical to keeping your agency out of the next headline, yet it often goes unevaluated. A 2025 report from the U.S. Government Accountability Office (GAO) found that fewer than half of all federal agencies evaluated safeguards for AI systems, with many having no process for assessing the trustworthiness of outputs.
Set yourself up for success by asking what safeguards are in place. Secure-by-design environments make sure that data only touches authorized infrastructure, is accessed only by individuals with a legitimate need, and is protected through zero-trust principles, detailed permissions, continuous monitoring, and comprehensive audit trails. Ultimately, building trust in AI outputs begins with trusting the environment that’s producing them.
2. How is the environment protected, and who is responsible for protecting it?
Technology alone won’t define whether agencies are ready for this new generation of AI-enabled threats. The most urgent investment is in operational readiness: building the proper governance frameworks and training the people responsible for managing the risk.
Advances in hardware and software are moving in sync. This convergence is creating what some experts refer to as a “technology tsunami.” Security assumptions that felt reasonable six months ago may already be outdated. The latest models and attack techniques are emerging so rapidly that agencies cannot solely rely on static policies to stay up to date.
For government leaders, the takeaway is clear: Be ready to unlearn fast. But it’s also critical to invest in governance structures that can adapt as quickly as the technology itself. If an agency has a Chief Information Security Officer, General Counsel, and a Chief AI Officer, who ultimately owns the security and defensibility of an AI-assisted workflow? Leverage a cross-functional governance body across legal, security, privacy, FOIA, and IT to manage AI approvals, policy updates, auditing requirements, and ongoing risk management.
At the same time, agencies must invest in their workforce. Building a culture of cybersecurity awareness and AI literacy is equally important as investing in the technology itself. Given leaner government teams, agencies should also look to partner agencies, contractors, or external experts to help fill in any critical capability gaps.
3. What third-party technologies are integrated into the workflow?
From cloud infrastructure to APIs and external data, AI environments rely on a greater ecosystem of third-party integrations. These dependencies introduce the greatest security risks.
For agencies, gaining a full picture of what third-party technologies are integrated into each workflow is essential. Vendors today need to effectively demonstrate a mature and evolving security program, one that continuously adapts to mitigate shifting risk.
Agencies should ask prospective vendors:
- What third-party products are integrated within your applications?
- How are these third-party technologies vetted or monitored?
- What happens when new models or software are upgraded?
- How do you continuously update protocols and processes?
4. What restrictions exist to limit access?
Whenever AI is embedded in FOIA and eDiscovery workflows, systems must be as closed as possible to minimize risk. Zero-trust principles should be the standard. The more open an environment is, the greater the attack surface and the greater the risk of exploiting sensitive information.
Additional tools should include IP blocking, country restrictions, and agency-only access controls to keep the potential for threats lower.
Agencies should also be mindful of “shadow AI,” where employees use unauthorized tools outside of established governance frameworks. Even well-intentioned experimentation can create significant security risk when sensitive government information is involved.
5. Can every AI-assisted action be audited and explained?
As AI shoulders a larger role in daily workflows, trust will depend on one critical question: Can agencies explain how each decision was made? If an AI agent modifies a workflow, changes a record, or deletes information, agencies must be able to pinpoint precisely what happened, why it happened, and who approved it.
Before introducing AI into an investigation, FOIA, or eDiscovery workflow, agencies should establish three foundational requirements:
- Provenance checks on every document entering the workflow
- Permanent AI-generated labeling on every output leaving the workflow
- Anomaly detection on the document population itself
No AI-generated output should enter the evidentiary record without a human reviewer verifying the source material and attesting that they reviewed the source and verified the output against it.
The Future of Security in FOIA Workflows
Agencies are rightly focused on AI productivity gains. But every productivity investment must be matched by an equal investment in security readiness. The cost of a security failure, compromised investigation, lost record, or improper disclosure will almost always exceed the productivity gains that led agencies to adopt the technology in the first place.
The agencies that succeed in this next era of AI-powered FOIA and eDiscovery will be the ones who center security, governance, and defensibility into every workflow from the very beginning.
Sundhar Rajan is Casepoint’s Chief Information Officer, responsible for overseeing information technology strategy, global infrastructure, and security compliances for the company to ensure that it meets client needs. He brings over 20 years of network engineering and infrastructure security experience. Prior to joining Casepoint, Sundhar was a senior network operations manager at a top Am Law 250 firm located in Washington D.C., where he led a team of 15 and enhanced network infrastructure security and applications deployment. Sundhar’s knowledge of information security, global infrastructure, and proactive compliance security monitoring aid him in leading Casepoint’s information technology.
Leave a Reply
You must be logged in to post a comment.