Companies & research organizations
Operational safeguards are being built before the policy debate is settled.
This directory tracks organizations developing controls that may help people monitor, constrain, interrupt, or evaluate advanced AI systems. Product descriptions are attributed to their sources; inclusion does not mean a safeguard has been independently proven or amounts to a universal kill switch.
Steven Jung, co-founder & CEO
Places controls inside agent execution environments to constrain risky actions before they occur.
Relevant work
Editorial scope: Its pre-execution controls are close to an operational agent kill switch, but remain product-layer safeguards.
Ankur Shah, co-founder & CEO; Sreenath Kurupati, co-founder & CTO
Discovers enterprise AI agents, tests them adversarially before deployment, and monitors activity at runtime.
Relevant work
Editorial scope: These are vendor-described controls for deployed agents, not evidence of a universal model-level shutdown mechanism.
Rick Caccia, co-founder & CEO
Observes and controls employee AI use and autonomous-agent behavior across enterprise environments.
Relevant work
Editorial scope: The company presents a governance and enforcement layer for enterprise agents, rather than a general-purpose frontier-model kill switch.
Ben Kliger, co-founder & CEO
Secures enterprise agent platforms against hijacking, misuse, and unauthorized autonomous actions.
Relevant work
Editorial scope: Zenity describes controls that can permit, block, or stop agent actions; effectiveness depends on deployment coverage and configuration.
Dario Amodei, co-founder & CEO
Uses filesystem, network, and execution boundaries to contain Claude across product environments.
Relevant work
Editorial scope: These controls are specific to Anthropic products and deployments; they do not establish a general external kill switch.
Matt Fredrikson, co-founder & CEO
Develops adversarially robust guardrails and crowdsourced testing for production AI systems.
Relevant work
Editorial scope: Red-teaming and guardrails improve detection and prevention; they are adjacent to, not substitutes for, reliable shutdown infrastructure.
Anand Kannappan, co-founder & CEO
Provides evaluation, guardrails, and scalable supervision intended to catch harmful agent behavior.
Relevant work
Editorial scope: Supervision can support intervention triggers, but it is not itself a guaranteed shutdown mechanism.
Pravin Kothari, founder & CEO
Combines agent identity, pre-execution policy checks, human approval, graduated containment, and circuit breakers.
Relevant work
Editorial scope: The company explicitly markets action blocking and termination controls; independent evidence of coverage is still limited.
Marc Benioff, chairman & CEO
Provides lifecycle management, guardrails, monitoring, and operational controls for deployed enterprise agents.
Relevant work
Editorial scope: Agentforce controls apply to agents managed within the platform, not to arbitrary AI systems outside it.
Demis Hassabis, CEO
Uses capability thresholds, safety cases, and protocols addressing models that might resist operator control or shutdown.
Relevant work
Editorial scope: This is a developer governance framework, not a commercial kill-switch product.
Beth Barnes, founder & CEO
Evaluates autonomy, self-replication, sabotage, and other agentic capabilities that may inform release gates.
Relevant work
Editorial scope: Evaluation results can inform no-deploy triggers, but the organization does not operate a kill switch.
Todd McKinnon, co-founder & CEO
Registers agent identities and can revoke credentials, permissions, and sessions at the identity layer.
Relevant work
Editorial scope: Identity controls can halt access to connected services but may not stop an agent operating with other credentials or infrastructure. Update Oct 15, 2026: an Okta-led alliance publicly urged an AI agent kill switch built on OAuth identity standards (ZDNet), and Okta shipped new agent-identity visibility features at Oktane 2026.
Jeffrey Ladish, executive director
Runs experiments on whether frontier models sabotage shutdown mechanisms, including in embodied settings.
Relevant work
Editorial scope: Palisade studies whether controls fail; it does not sell a shutdown product.
Nikesh Arora, chairman & CEO
Offers an enterprise AI control plane and runtime security protections for models, applications, and agents.
Relevant work
Editorial scope: These enforcement boundaries are enterprise security controls, not independently verified universal shutdown systems.
Marius Hobbhahn, co-founder & CEO
Builds evaluations and monitoring methods for deceptive or scheming frontier models.
Relevant work
Editorial scope: Detection supports intervention decisions, but reliable detection and complete containment remain open problems.
Current program leadership not independently verified
Builds security-focused reasoning models and agentic tools for cybersecurity workflows.
Relevant work
Editorial scope: Yaron Singer led the program when the cited work launched but reportedly left Cisco in July 2026; current leadership is not assigned here without confirmation.
Buck Shlegeris, CEO
Develops monitoring and control protocols intended to preserve operator control over potentially subversive models.
Relevant work
Editorial scope: The work is research on control and shutdown-safe scaffolding rather than a deployed commercial switch.
Roy Katmor, co-founder & CEO; Robert Wiseman and Ido Kelson, co-founders
Uses language-model reasoning to discover applications, map identities and credentials, and support access revocation.
Relevant work
Editorial scope: This is adjacent identity governance, not dedicated AI-agent runtime containment or a general-purpose kill switch.