Open Problems in AI Governance
We publish the hardest unsolved questions in AI governance, causal AI, and policy-constrained systems. Solve one and become a Chimera Fellow.
Can formal policy constraints prevent prompt injection attacks?
Investigate whether CSL-Core constraints can define input boundaries that make prompt injection structurally impossible. Explore how constraint verification at the policy layer differs from output filtering approaches.
How can AI agents produce verifiable compliance proofs?
Design a mechanism where AI agents generate cryptographic or logical proofs that their actions comply with a given CSL policy set. The proof should be independently verifiable without re-running the agent.
Can causal traces improve explainability of AI decisions?
Explore whether formal causal traces (as defined in CSL-Core) can serve as human-readable explanations for AI decision-making. Compare causal trace explainability against SHAP, LIME, and attention-based methods.
What is the minimal constraint system for safe autonomous agents?
Define the smallest set of CSL constraints that guarantee safety for a general-purpose autonomous agent. Investigate whether there exists a universal safety kernel that all agent architectures must satisfy.
Benchmark: LLM guardrails vs CSL formal constraints
Create a rigorous benchmark comparing traditional LLM guardrails (Constitutional AI, RLHF, output filters) against CSL-Core formal constraints. Measure violation rates, latency overhead, and coverage across adversarial scenarios.
Can CSL policies compose across multi-agent systems?
Investigate policy composition: when multiple agents each follow individual CSL policies, does the composed system satisfy a global safety property? Formalize conditions under which composition is safe.
SUGGEST A PROBLEM
Have an open question?
Submit a research question. If accepted, it gets published here with credit to you.
Think you can solve one? Join the Research Program.
Successful contributors become Chimera Fellows with recognition, portfolio credit, and ongoing research access.