Two Programmes, One Problem
What DARPA's CLARA and SABER Programmes Reveal About Governing Autonomous AI at the Edge
Prove It, Then Break It
What DARPA's CLARA and SABER Programmes Reveal About Governing Autonomous AI at the Edge
Reference: EA/DEF/2026/002 · Date: 22 October 2026 · Series: Defence Technology & AI Governance
• • •
Two Programmes, One Problem
In 2025 and 2026, the US Defense Advanced Research Projects Agency (DARPA) launched two AI programmes that, read together, describe the central problem of autonomous systems more clearly than most policy documents do.
The first, CLARA, asks how to build AI whose behaviour can be proven. The second, SABER, asks how AI that is already fielded can be broken. One programme works on assurance from the inside, through mathematics and logic. The other works on assurance from the outside, through adversarial testing on a representative battlefield.
Neither alone is sufficient. A proof only holds over the model of the world it was written for, and an attacker's job is to find the places where the world departs from that model. A red team, in turn, can only demonstrate that it failed to break a system; it cannot demonstrate that the system cannot be broken. Governance of autonomous AI lives in the gap between the two.
• • •
CLARA: Building AI That Can Be Proven
| Item | Detail |
|---|---|
| Full name | Compositional Learning-And-Reasoning for AI Complex Systems Engineering |
| Office | Defense Sciences Office (DSO) |
| Programme manager | Benjamin Grosof |
| Funding | Up to US$2 million per project, as Other Transaction prototype agreements |
| Structure | 15-month Phase 1 base; 9-month Phase 2 option |
| Proposal deadline | 17 April 2026 |
| Technical Area 1 | New high-assurance ML/AR theory, algorithms, and open-source software for inference and training |
| Technical Area 2 | A software integration library: common data formats, interfaces, end-to-end explainability tools |
The core idea
Most current approaches to trustworthy AI bolt a reasoning component onto a large language model as a safeguard. DARPA's own description is pointed: the common practice is to "tack specialized automated reasoning (AR) components onto a large language model."
CLARA proposes the opposite. Machine learning (ML) and automated reasoning are composed together from the start, in a "hierarchical, fine-grained, highly transparent" structure. Bayesian models, neural networks and logic programs become interchangeable building blocks, and the target is "verifiability based on AR proofs that have strong logical explainability and computational tractability."
The listed application areas are revealing: kill webs, logistics, wargaming, autonomous and command-and-control systems, but also medical, financial and legal domains. CLARA is not a weapons programme. It is a bet that high-assurance AI is a general engineering discipline, in the same way that structural analysis is general.
• • •
SABER: Breaking AI Before the Adversary Does
| Item | Detail |
|---|---|
| Full name | Securing Artificial Intelligence for Battlefield Effective Robustness |
| Programme manager | Nathaniel Bastian |
| Target systems | AI-enabled autonomous ground and aerial systems that could be fielded within one to three years |
| Proposers Day | 12 March 2025 |
| Threat classes | Data poisoning, physically constrained adversarial patches, model stealing |
| Techniques in scope | Physical, adversarial-AI, cybersecurity and electronic warfare methods |
Known awards
| Performer | Value | Reported role |
|---|---|---|
| BAE Systems Electronic Systems | US$3.9 million (early February 2026) | Counter-AI techniques and tools for operational assessment |
| Lockheed Martin Advanced Technology Laboratories | US$4.7 million (announced February 2026) | Counter-AI techniques and tools for operational assessment |
| Kitware, with Scientific Systems Company | US$4 million, two years (May 2026) | Physical-world adversarial attacks, evaluation tools, field testing on ground and rotary-wing systems |
The core idea
SABER's starting observation is that adversarial attacks on AI have mostly been demonstrated in simulation or in physical settings with "relatively pristine control." Almost nobody has shown what happens when a poisoned dataset, a printed adversarial patch or a stolen model meets a real autonomous vehicle in a contested environment. DARPA's stated position is that no current methods exist to assess deployed military AI systems for these vulnerabilities.
The programme's deliverable is therefore not a single tool, but a repeatable model for operational AI red teaming: people, methods and test procedures that can be applied to each new system before it is trusted.
• • •
Why the Edge Changes Everything
Governing AI in a data centre is a policy problem. Governing AI at the edge is a control problem. Four conditions make it so:
No reliable link: An edge system may lose contact with its operators, exactly as a hypersonic vehicle loses GNSS in its plasma sheath. Governance that depends on a remote human approving each decision fails at the moment it is needed.
Tight resource limits: Size, weight, power and cost constraints rule out running large verification engines onboard. Any runtime check must be cheap.
Adversarial inputs: The sensors are pointed at the world, and the world can be edited. A patch on a vehicle or a pattern on a road is an attack on the perception stack that requires no network access at all.
Physical consequences: A wrong output is not a bad answer on a screen. It is a vehicle in the wrong place.
These conditions explain why the two programmes fit together. CLARA addresses what can be guaranteed onboard with limited compute. SABER addresses what the adversary will do to the inputs those guarantees depend on.
• • •
The AMANAH Lens: An Islamic Governance Ontology for Autonomy
The AMANAH framework approaches the same problem from a different direction. It draws on three concepts from the Islamic governance tradition and gives each an engineering interpretation.
Amanah: bounded, entrusted authority
Amanah is a trust held on behalf of another. The one who holds it is accountable for it and may not exceed it. For an autonomous system, this becomes a formal operating envelope: a set of states and actions within which the system is authorised to act alone, and outside of which it must defer, degrade or stop. Authority is delegated, never owned.
Muraqabah: continuous self-observation
Muraqabah is watchfulness, the awareness of being observed and the discipline of observing oneself. In engineering terms, this is a runtime monitor that runs continuously alongside the decision-making component and checks whether the system is still inside its envelope, and whether its inputs still look like the world it was trained on.
Adab: norms of conduct
Adab is right conduct: the rules of behaviour that hold regardless of the immediate goal. In engineering terms, these are explicit, inspectable constraints, which is precisely the role CLARA assigns to automated reasoning components composed alongside learned ones.
Mapping the frameworks
| AMANAH principle | Engineering mechanism | CLARA contribution | SABER contribution |
|---|---|---|---|
| Amanah (bounded authority) | Safety envelope defined by a Lyapunov-type function; authority limited to the region where it is guaranteed to decrease | Proofs that the composed system stays within the envelope | Tests of whether adversaries can push the system across the boundary |
| Muraqabah (self-observation) | Runtime assurance monitor; input-distribution checks | Tractable reasoning that can run in real time | Evidence of which attacks evade the monitor |
| Adab (norms of conduct) | Explicit logical constraints composed with learned components | Hierarchical ML/AR composition with explainability | Validation that constraints hold under poisoned or manipulated inputs |
A concrete architecture
The mapping above corresponds to a well-established pattern in safety-critical control, often called the Simplex or runtime-assurance architecture:
An advanced controller (which may be a learned policy) handles normal operation and delivers high performance.
A verified safety controller, simple enough to be proven correct, can always return the system to a safe state.
A decision module monitors the state and hands control to the safety controller whenever the system approaches the edge of its envelope.
In AMANAH terms, the advanced controller exercises entrusted authority, the decision module performs muraqabah, and the safety controller embodies adab: the conduct that holds when everything else is uncertain. The Lyapunov function defines where the trust ends.
• • •
What This Means for Researchers Outside the United States
CLARA and SABER are US defence programmes, but their outputs and their lessons travel:
Open-source software is coming. CLARA's Technical Area 1 explicitly includes open-source software for inference and training. Research groups anywhere will be able to build on, benchmark against and critique it.
The red-team method is the transferable asset. SABER's lasting contribution is likely to be a methodology, not a weapon. A university UAV group can adopt the same discipline today: define the threat classes, attack your own perception stack physically, and publish what breaks.
Governance frameworks need engineering hooks. Ethical principles for AI are abundant. Principles that compile down to an envelope, a monitor and a constraint set are rare. Frameworks that make that connection explicit, from any intellectual tradition, have something to offer the field.
Civilian autonomy is under-tested. The attack surface SABER studies is already present in commercial drones and robotics. Independent, transparent red-teaming of civilian systems is an open niche.
• • •
Conclusion
CLARA and SABER are best read as two halves of one sentence: prove what you can, then attack what you proved. The proof tells you the boundary of trust. The attack tells you whether the boundary holds in the real world.
The AMANAH framing adds a third half that engineering alone tends to leave implicit: that autonomy is a delegated trust, not an inherent right. A system that knows the limits of its authority, watches itself continuously, and holds to fixed norms of conduct when uncertain is not merely safer. It is governable. That, rather than raw capability, is the property that will decide which autonomous systems are allowed to operate at all.
• • •
إِلٰهِي أَنْتَ مَقْصُودِي وَرِضَاكَ مَطْلُوبِي، أَعْطِنِي مَحَبَّتَكَ وَمَعْرِفَتَكَ
Ilahi anta maqsudi wa ridhaka matlubi, a'tini mahabbataka wa ma'rifataka
Prepared by TXIO Fusion Solutions
Disclaimer: This document is prepared for research awareness purposes only. It is based on open-source reporting available as of the publication date, including official DARPA programme pages and published contract announcements, and may not reflect subsequent developments. It does not constitute policy, legal or investment advice. TXIO Fusion Solutions assumes no liability for decisions made on the basis of this document.
Sources
DARPA, "CLARA: Compositional Learning-And-Reasoning for AI Complex Systems Engineering" (programme page)
DARPA, "SABER: Securing Artificial Intelligence for Battlefield Effective Robustness" (programme page)
Military Embedded Systems, "AI assurance the aim of new DARPA program" (11 February 2026)
Military & Aerospace Electronics, "DARPA seeks high-assurance artificial intelligence proposals under CLARA program" (13 April 2026)
Military & Aerospace Electronics, "Lockheed Martin, BAE to defend military AI from data poisoning and cyber attack" (12 March 2026)
Kitware, "Kitware Awarded $4M DARPA SABER Contract to Evaluate AI-Enabled Battlefield Systems" (4 May 2026)
EverGlade Consulting, "DARPA Launches SABER Initiative to Secure AI on the Battlefield" (17 February 2025)