The Agent Stack: Security & Control
The OpenAI and Hugging Face incident showed that autonomous agents cannot be trusted to report or police their own behavior. As agents gain more authority, companies will need an independent control and evidence layer around them. This landscape focuses on new pure play companies building that layer rather than established security platforms expanding into agent security.
Explore
Categories
- Tool & MCP Security
Agents increasingly rely on external tools to act. These products provide an independent control layer that decides which MCP servers or tools an agent is allowed to use.
- Agent Runtime & Execution Security
Approved tools can still be used in unsafe ways. These products provide an independent control layer around execution and stop actions that break policy before they affect external systems.
- Agent Testing & Assurance
Some agent failures only appear when the system is pushed into difficult situations. These products test agents independently to expose dangerous behavior before it reaches production.
- Agent Observability & Evidence
Control is not enough if the evidence comes from the agent itself. These products create an independent record of execution so teams can verify what the agent actually did.
- Agent Identity & Authorization
Infrastructure that establishes the identity and provenance of agents and determines what they are permitted to access or do before execution.
Companies
- Agnys
- AIM Intelligence
- AIR
- AI Score
- Anomity
- Asymptote Labs
- Bay
- Beltic
- Capsule Security
- CodeIntegrity
- Cygrid
- Eve Security
- Evoke Security
- Fact0
- Giskard
- Goodfire
- Gray Swan
- Harbor
- Harmonic Security
- HiddenLayer
- Kontext
- Lasso Security
- Lumbra
- Manifold
- Mintive
- NeuralTrust
- NOFire
- Noma Security
- Operant AI
- Opsin
- Outerlimit
- Paper Compute
- Pillar Security
- Pluto Security
- Quint Security
- Raindrop
- Rein Security
- Runlayer
- Straiker
- Tenet AI
- Trent AI
- Ultra Security
- Vaara
- Verno Labs
- WhiteMatter
- WitnessAI