The UC Berkeley Center for Long-Term Cybersecurity (CLTC) is pleased to announce the 8th annual Cal Cybersecurity Research Fellowship. The 2026-27 fellowship supports innovative research at the intersection of cybersecurity and artificial intelligence, with a focus on emerging threats posed by and to generative AI and agentic systems. The successful proposal will receive an award of up to $15,000.
We invite proposals from UC Berkeley graduate students, post-doctoral scholars and pre-tenure faculty pursuing novel research in the priority areas described below. While these priority areas represent topics of special interest, strong proposals that fall outside the priority areas but demonstrate significant potential for novel scientific advances in AI-cybersecurity will receive consideration.
Priority Research Areas
1. Cognitive Firewalls: Defending AI Reasoning Against Manipulation
Prompt injection combined with tool access represents one of the most dangerous practical risks with generative AI today. As AI agents increasingly operate with autonomy—reading emails, browsing the web, invoking tools, and communicating with other agents—their reasoning processes become attack surfaces.
Today’s firewalls protect networks. Future systems may require firewalls that protect reasoning.
A cognitive firewall would defend AI reasoning by identifying threats such as:
– Prompt injection
– Authority confusion
– Manipulation attempts
– Social engineering
– Conflicting objectives
These threats may arrive through multiple channels, including:
– User prompts
– Retrieved documents
– Emails and web pages
– Inter-agent messages
– Tool outputs
We are seeking proposals that advance defenses against these threats through any promising approach, including but not limited to:
– Inspecting and filtering inputs to AI models
– Model-level defenses (e.g., post-training techniques to improve resilience)
– Systems-level and architectural defenses (e.g., new agent harnesses, sandboxing, privilege separation)
– Novel agent architectures designed to resist manipulation
– Hybrid approaches combining multiple layers of defense
A strong proposal does not need to address all of the threats listed above. We welcome proposals that articulate a persuasive direction to make novel scientific advances on one or more of these threats.
We also welcome proposals that address fundamental challenges in this space, such as how detection mechanisms themselves can be made robust against the same attacks they aim to prevent.
Key questions include:
– How can cognitive defense capabilities be designed and implemented?
– What architectures make AI reasoning robust against adversarial inputs?
– How can we verify that defense mechanisms are not themselves vulnerable to manipulation?
2. Computational Trust Models for AI Agents
Today’s large language models largely treat all text as semantically equivalent. Humans, by contrast, evaluate information based on source identity, authority, expertise, provenance, consistency, incentives, and historical reliability.
Can an AI agent quantify how much it should trust every piece of information it receives?
Research in this area would develop formal trust models where every observation carries attributes such as:
– Source identity
– Confidence
– Authority
– Freshness
– Integrity
– Provenance
– AI integrity (e.g., whether the information originates from a verified, uncompromised and aligned model)
– Historical reliability
The agent would reason over trust rather than simply reasoning over language. This may resemble zero-trust networking, but applied to cognition — including zero trust of AI systems themselves.
This research area is inherently interdisciplinary. We welcome proposals approaching computational trust from multiple angles, including:
- Sociotechnical approaches: Who gets to define trustworthiness? How is trust contextual (e.g., a source may be trusted in one domain but not another)? What frameworks from library science, information science, and epistemology apply?
- Technical approaches: Can interpretability techniques reveal how trust is encoded in model representations? Could trust assessments be read from the residual stream of open-weight models? How can trust scores be computed and propagated in multi-agent systems?
- Formal/mathematical approaches: What formal models of trust (e.g., from distributed systems, reputation systems, or Bayesian frameworks) can be adapted for AI agents?
Key questions include:
– How should trust be represented, computed, and updated in AI systems?
– How can contextual trust be modeled (trusting a source for some purposes but not others)?
– How can trust models be made robust against adversarial manipulation of trust signals?
Eligibility
This award may be used only to support a current graduate student, postdoc, or pre-tenure faculty member’s research efforts. The award funds are discretionary and can be applied to any research expenses, including but not limited to: student/postdoc salary on an official university appointment, conference registration and travel, research supplies, and other direct costs related to the project.
You can see examples of previously awarded Cal Cybersecurity Research Fellowships here.
How to Apply
If you would like to apply for the Research Fellowship, please submit the following information to this form by October 16, 2026 at 11:59 PM PT:
- CV
- Project proposal (5-page maximum) including a project title, abstract, previous related work, research methods, and expected outcomes and impact.
- Appendix: append a brief biography of you and your advisor(s) and a simple budget describing how you would use the award funds.
- Please note that CV and appendix do not count toward the 5-page limit.
Questions may be directed to Audrie Hough at cltcgrants@berkeley.edu
About CLTC
The Center for Long-Term Cybersecurity (CLTC) is a premier hub for research and collaboration on public-interest cyber and AI security. Housed within the University of California, Berkeley and affiliated with the School of Information, CLTC looks “over the horizon” to anticipate and address tomorrow’s digital security challenges.
Our community of students, researchers, and practitioners works to protect our shared digital life. We believe that navigating a volatile digital landscape requires more than just technology; it requires a persistent, human-centered approach to shape a secure future for everyone.
