AIR Security is an Agentic AI Security vendor based in New York, USA; staff in Israel. Its flagship product is AIR (AIR Control, AIR Filter, AIR Defend, AIR Marketplace). Against the OWASP LLM Top 10 (2026) we score it Covered on 0 of 10 risks, Partial on 3, with 4 MITRE ATLAS techniques mapped as of . It's independent.
- 0/10
- Risks covered
- 3
- Partial
- 4
- ATLAS techniques
- $50M
- Funding
- Active
- Status
At a glance
- Category
- Agentic AI Security
- Headquarters
- New York, USA; staff in Israel
- Founded
- 2026
- Employees
- 11-50
- Flagship product
- AIR (AIR Control, AIR Filter, AIR Defend, AIR Marketplace)
- Deployment
- Sold through AWS Marketplace and Google Cloud Marketplace; deployment architecture is not publicly documented as of 2026-10-30
- Best for
- Enterprises with heavy coding-agent use that need continuous vetting of the skills, plugins and MCP servers those agents install
- Funding to date
- $50M
- Last round
- Seed ($40M) · 2026-09
- Status
- Active
Facts as of .
What does AIR Security do?
A firewall for agent context: vets skills, plugins, MCP servers and sub-agents before they reach an agent, then rechecks them after updates. It ships as Sold through AWS Marketplace and Google Cloud Marketplace; deployment architecture is not publicly documented as of 2026-10-30, and it's built for enterprises with heavy coding-agent use that need continuous vetting of the skills, plugins and MCP servers those agents install. It's a weaker fit for buyers who need documented model-level guardrails, data protection, or a public technical reference before a demo.
Which OWASP LLM Top 10 risks does AIR Security cover?
The two tables below come from our editorial read of AIR Security's public documentation as of (it's what the docs support, not a hands-on lab test), and this row also sits in the full matrix across all 27 vendors.
| OWASP LLM Risk | What it is | AIR Security coverage | How it's addressed | Source |
|---|---|---|---|---|
| LLM01 Prompt Injection | User or hidden input overrides the model's rules or intended behavior. | Partial | AIR catches prompt injection in add-ons before production, but no runtime control over direct user prompts is documented. | https://www.air.security/blog-posts/air-and-google-cloud-marketplace |
| LLM02 Sensitive Information Disclosure | The model leaks secrets, personal data, or proprietary content in its output. | Unverified | AIR lists data exfiltration as an add-on risk, but no data classification, redaction or DLP control is documented. | https://www.air.security/ |
| LLM03 Excessive Agency | An agent holds more permissions, tools, or autonomy than the task needs. | Partial | AIR Control holds every agent to policy on config, identity and permissions, but runtime enforcement detail is not published. | https://www.air.security/ |
| LLM04 Supply Chain | Compromised models, datasets, plugins, or dependencies add risk before runtime. | Partial | AIR Filter scans add-on dependencies and sandboxes skills, but no model-file or dataset provenance control is documented. | https://www.air.security/blog-posts/anthropic-scanner |
| LLM05 Data and Model Poisoning | Tampered training or fine-tuning data corrupts how the model behaves. | Not covered | AIR's documented scope is agent add-ons and runtime behavior, not training data or fine-tuning. | https://www.air.security/blog-posts/out-of-stealth |
| LLM06 Unbounded Consumption | Uncontrolled requests drive cost, denial of service, or model extraction. | Unverified | No rate limit, token budget or cost control found in AIR material as of 2026-10-30. | None found |
| LLM07 Misinformation | The model produces false or fabricated content that users act on. | Unverified | No hallucination or factual-accuracy control found in AIR material as of 2026-10-30. | None found |
| LLM08 Hidden Context Exposure | Hidden instructions, retrieved policy text, or tool schemas get extracted, inferred, or reconstructed from the model. | Unverified | No system-prompt or hidden-context protection found in AIR material as of 2026-10-30. | None found |
| LLM09 Vector and Embedding Weaknesses | Flaws in RAG stores let attackers poison, extract, or infer data. | Unverified | No vector-store or embedding control found in AIR material as of 2026-10-30. | None found |
| LLM10 Improper Output Handling | Downstream systems trust model output without checking it, enabling injection or code execution. | Unverified | No output validation control found in AIR material as of 2026-10-30. | None found |
| ATLAS Technique (ID) | Tactic | AIR Security coverage | Notes | Source |
|---|---|---|---|---|
| LLM Prompt Injection (AML.T0051) | Execution | Partial | AIR catches prompt injection carried in add-ons before production. | https://www.air.security/blog-posts/air-and-google-cloud-marketplace |
| AI Agent Tool Invocation (AML.T0053) | Execution / Privilege Escalation / Lateral Movement | Unverified | AIR says its runtime layer enforces agent actions, but names no tool-call control. | https://www.air.security/blog-posts/claude-air-integration |
| AI Agent Tool Poisoning (AML.T0110) | Persistence | Covered | AIR Filter sandboxes skills and scans dependencies to expose hidden behavior in agent add-ons. | https://www.air.security/blog-posts/anthropic-scanner |
| Supply Chain Compromise (AML.T0010) | Initial Access | Partial | AIR Filter and AIR Marketplace vet agent add-ons, but no control for models or datasets is documented. | https://www.air.security/blog-posts/plugin4shell |
Which OWASP Agentic risks does AIR Security cover?
1 of 10 Covered and 3 Partial against the OWASP Top 10 for Agentic Applications, scored from public documentation as of 2026-10-30.
| Agentic Risk | What it is | AIR Security coverage | How it's addressed | Source |
|---|---|---|---|---|
| Agent Goal Hijack (ASI01) | Attackers redirect what an agent is trying to do, often through documents or other content it reads. | Partial | AIR filters untrusted inputs before they reach agent context, with add-ons the documented path, so hijack through other content is unconfirmed. | https://www.air.security/blog-posts/out-of-stealth |
| Tool Misuse and Exploitation (ASI02) | An agent uses a legitimate tool in a harmful way, or a manipulated tool call does damage. | Unverified | AIR says its runtime layer enforces agent actions one by one, but no tool-level control is described. | https://www.air.security/blog-posts/claude-air-integration |
| Identity and Privilege Abuse (ASI03) | Agents inherit, reuse, or escalate credentials and permissions beyond what the task needs. | Partial | AIR identifies risky permissions across the agent fleet, but enforcement detail is not published. | https://www.air.security/blog-posts/air-and-aws-marketplace |
| Agentic Supply Chain Vulnerabilities (ASI04) | Compromised MCP servers, plugins, skills, or agents enter the runtime. | Covered | AIR Filter vets skills, plugins, MCPs and subagents before install, and AIR Marketplace supplies pre-vetted add-ons. | https://www.air.security/ |
| Unexpected Code Execution (ASI05) | An agent generates or runs code that gives an attacker execution. | Partial | AIR Filter sandboxes skills to find hidden behavior before use, but blocking code execution at runtime is not documented. | https://www.air.security/blog-posts/anthropic-scanner |
| Memory and Context Poisoning (ASI06) | Attackers plant content in an agent's memory or context that shapes later decisions. | Unverified | AIR describes filtering what enters agent context, but no control for persistent agent memory is documented. | https://www.air.security/blog-posts/out-of-stealth |
| Insecure Inter-Agent Communication (ASI07) | Messages between agents get spoofed, intercepted, or tampered with. | Unverified | No inter-agent message authentication or integrity control found in AIR material as of 2026-10-30. | None found |
| Cascading Failures (ASI08) | One fault spreads across connected agents and systems. | Unverified | No containment control for failures spreading across agents found in AIR material as of 2026-10-30. | None found |
| Human-Agent Trust Exploitation (ASI09) | An agent's confident output talks a person into approving something harmful. | Unverified | No control for manipulated human approval found in AIR material as of 2026-10-30. | None found |
| Rogue Agents (ASI10) | An agent drifts or is compromised and acts outside policy. | Unverified | AIR Defend is described as detect, respond and protect in real time, but no rogue-agent detection method is documented. | https://www.air.security/ |
Is AIR Security independent, and how is it funded?
AIR Security is an independent company as of . It has raised $50M to date, most recently a Seed ($40M) dated 2026-09. Lead investors: Sequoia Capital, Greenoaks.
AIR Security alternatives
The closest alternatives we track in Agentic AI Security are Akto, Knostic, Straiker. On the open-source side, agent-scan covers similar ground.
Frequently asked questions
What is AIR Security used for?
AIR Security is an Agentic AI Security vendor. Its flagship product is AIR (AIR Control, AIR Filter, AIR Defend, AIR Marketplace). It ships as Sold through AWS Marketplace and Google Cloud Marketplace; deployment architecture is not publicly documented as of 2026-10-30, and it's built for enterprises with heavy coding-agent use that need continuous vetting of the skills, plugins and MCP servers those agents install.
Is AIR Security independent or acquired?
AIR Security is an independent company as of 2026-10-30 and has not been acquired.
How many OWASP LLM Top 10 risks does AIR Security cover?
We score AIR Security as Covered on 0 of the ten OWASP LLM Top 10 (2026) risks, with 3 more Partial, as of 2026-10-30. The table on this page breaks down every risk, including the ones marked Not covered or Unverified.
How is AIR Security deployed?
AIR Security ships as Sold through AWS Marketplace and Google Cloud Marketplace; deployment architecture is not publicly documented as of 2026-10-30.
What are the best alternatives to AIR Security?
The closest Agentic AI Security alternatives llmthreat tracks are Akto, Knostic, Straiker. On the open-source side, agent-scan covers similar ground.