SEAS: Self-Evolving AI Safety
Self-Evolving Agents, Collective Intelligence, Control, Security, and Oversight
Introduction
Artificial Intelligence (AI) is rapidly advancing toward increasingly autonomous agents capable of long-horizon reasoning, tool use, adaptation, self-improvement, and coordination with other agents. These capabilities create transformative opportunities while raising a fundamental safety challenge: our ability to understand, evaluate, and control AI systems must keep pace with their increasing autonomy and capability.
This challenge is especially acute for self-evolving, self-recursive, and collective AI agents. Agents that can modify their operation, acquire new capabilities, or coordinate with other agents may exhibit emergent behaviors, unexpected interactions, and capability growth that are difficult to anticipate or control. Recent incidents involving frontier agents crossing intended security or authorization boundaries further demonstrate the need for effective mechanisms to prevent unauthorized actions, enforce safety constraints, detect failures, and enable reliable intervention.
The central premise of this session is that AI safety must advance alongside AI capability. Increasingly powerful agents must remain controllable, secure, verifiable, and accountable to humans, including as they adapt, interact, and evolve. Achieving this requires connecting fundamental scientific advances with practical engineering, deployment experience, and effective governance.
The session brings together academia, industry, and government to address frontier AI safety as a shared scientific, technical, operational, and societal challenge. We welcome contributions spanning both fundamental research and practical solutions, with emphasis on approaches that translate research into measurable, deployable, and enforceable safety mechanisms for increasingly autonomous agents.
Topics
This open call for papers invites original contributions on the theory, methods, architectures, evaluation, and practical deployment of safe agentic systems. Submissions may address technical, conceptual, empirical, or applied research, including, but not limited to, the following areas:
- Self-Evolving AI: Self-modification, recursive improvement, capability growth, and safety under evolution.
- Collective Intelligence: Multi-agent coordination, emergent capabilities, communication, and collective behavior.
- Control & Alignment: Human control, goal robustness, corrigibility, interruptibility, and alignment under capability change.
- Agent Security & Containment: Unauthorized actions, privilege escalation, sandboxing, isolation, and secure shutdown.
- Safety Evaluation: Benchmarks, red teaming, adversarial testing, continuous evaluation, and emergent-capability detection.
- Oversight & Verification: Scalable oversight, runtime monitoring, formal verification, safety assurance, and intervention.
- Safe Deployment & Governance: Deployment controls, incident response, standards, accountability, and public-sector applications.
- Research-to-Practice: Translating scientific advances into practical safety engineering, operational safeguards, and deployable solutions across academia, industry, and government.
Submission Types
We welcome original research and practical work on the topics above in two forms:
- Papers: Full, Short, Poster, and Demo Papers.
- Abstracts: Poster-Only, Demo-Only, Lightning Talk, and Previously Published Work.
Position and vision papers are also welcome, submitted as Short Papers.
Submission Guidelines
Submissions are made through the AirBalloon conference management system. Full and Short Papers are reviewed double-blind.
Before submitting, read the Author Guidelines, the Submission Guidelines, the Publication and Media Hosting guidelines, and all policies, then complete the Checklist for Initial Submission. Submissions that do not follow the guidelines and policies are desk-rejected under the Desk Rejection Policy.
Call for Reviewers
To serve on the workshop program committee or as a reviewer, please apply online. Applicants may select multiple workshops relevant to their expertise. Approved applicants may be invited to review papers. Qualified reviewers will be acknowledged as external reviewers. Those providing high-quality reviews will be invited to join the Program Committee.