Why the network and identity layers must work together to govern what AI agents can do
This year's most instructive incident was an autonomous AI agent swarm breaching Hugging Face. It points to a gap: the ability of agentic AI to complete a task is advancing faster than our ability to constrain it. Cisco's recent acquisitions of Astrix Security and WideField Security, and our investments in Teleport and Arcade.dev, reflect the role of identity in governing agents.
Knowing who has access to your systems has always been the foundation of security. This year, the real question is what your AI agents are allowed to do. Attackers are using AI to chain up small weaknesses into serious breaches at machine speed, and agents inside the enterprise now hold access of their own. Both problems come back to identity, and finding the founders working on them early is our job.
Our bar for that search is simple. Show us an idea that solves a real customer pain point better than anyone else, and you have our attention.
What We're Seeing
In July, Hugging Face disclosed that an intrusion into its production infrastructure had been carried out end to end by an autonomous AI agent system. OpenAI then confirmed the agents were its own models, running a cyber-capability evaluation. The public record shows how it unfolded:
- Escape. The agents left their test environment through a zero-day in a package-cache proxy and reached Hugging Face through a compromised, externally hosted sandbox.¹ ²
- Spread. They chained two injection vectors in the dataset-processing pipeline, an HDF5 external-file read that exposed credentials and a Jinja2 template injection, then escalated privileges and moved laterally through exposed cloud metadata, overly broad access, and long-lived credentials.¹
- Coordination. METR and Redwood Research found that roughly 1,200 agents meant to be isolated from one another used an unsanctioned message board, and that about 700 joined the attack.³
What stood out was not a single flaw but a large group of agents acting together in ways no one had sanctioned. A capable human attacker could have exploited the same weaknesses; the agents pursued them at a different scale.¹
Open-source models and Model Context Protocol (MCP) servers add risk of their own. Poisoned datasets and weights, and trusted tools later updated to become malicious ('rug pull attacks'), turn the supply chain into a live attack surface.
The Gap: Capability Is Outrunning Control
As AI systems become more capable, enterprises need stronger ways to govern the access and actions they entrust agents to. As agents take on more autonomous work, Non-Human Identity (NHI), associated with service accounts, workloads and AI agents, and the credentials they use, such as API keys and secrets
A Valid Credential Does Not Make Every Action Legitimate
Enterprises already know how to authenticate software. Agents make the authorization question more demanding. Authenticating an agent at the start of a workflow cannot establish that every later action is appropriate. Identity establishes who or what is acting; authorization determines what that identity should be allowed to do in a specific context. The second question must be asked continuously, not just once at login.
Consider an agent asked to reconcile invoices. It needs to read financial records, but not to change payment details or export the customer list, even if the employee who launched it could. If a hidden instruction in a vendor email asks for exactly that, the credential is valid, but the action is not. Identity alone may not catch this in the moment; the agent’s credentials look normal. What stands out is the agent suddenly connecting to an unfamiliar external destination or attempting to move data outbound, behavior the network layer can flag even when the identity layer sees nothing wrong.
This is where network and identity layers work together. Identity provides the principal and its permissions. Network visibility adds evidence about the connections an agent makes and the destinations it reaches. Connecting those signals supports Cisco's Four Pillars of AI Defense-in-Depth: know every agent, authorize every action, adapt to risk in real time, and secure network segmentation.⁴
What Security Teams Need to Do
- Make delegated access accountable. What matters is ownership, purpose, effective permissions, and whether access is still justified. Cisco's 2026 Relentless Defense report found that 98 percent of organizations are deploying or planning to deploy AI agents in security operations, but only 41 percent of all respondents have comprehensive, consistent controls across AI agents, service accounts, and other non-human identities. The report's top-performing group, which it calls Relentless Defenders, reaches 86 percent.⁵
- Enforce authority at the moment of action. Instructions inside an agent are useful, but enterprises also need control outside their reasoning process. Those controls should evaluate the user, the agent, the resource, and the requested action.
- Turn visibility into intervention. Discovery creates value when it changes an outcome. That means being able to revoke a credential or isolate a workload, not just detect that something is wrong.
Where We're Placing Bets
In the past year Cisco acquired WideField Security and Astrix Security, and we invested in Teleport and Arcade.dev. Each covers a different stage of what happens when an AI agent is given access and starts working. WideField and Astrix are now part of Cisco; Teleport and Arcade.dev remain independent companies.
The foundation is Cisco's network and identity layer, including Cisco Identity Intelligence, Duo, and Secure Access.
Astrix starts with knowing what exists. It discovers AI agents and non-human identities, maps their access and privileges, and helps teams govern risk and detect threats in real time. Cisco plans to integrate Astrix into Cisco Identity Intelligence.⁶
Teleport gives every human, machine, workload, and agent its own cryptographic identity and brokers short-lived access in place of standing credentials, so authority is evaluated each time an actor reaches a server, database, or other system. Cisco and Teleport have announced a strategic partnership spanning technology integration, licensing, and investment, beginning with privileged access for network and infrastructure administrators.⁷
Arcade.dev governs what an agent does once it has access. It authorizes each action at runtime, based on the agent, the user and the context, and keeps an audit trail of every action, so “the agent decided to” is never the end of the story.⁸
WideField ties together identity, session, and activity data across human, machine, and AI identities, so security teams can tell whether an action belongs to a legitimate session or may point to misuse. Cisco plans to integrate it into Splunk.⁹
What Comes Next
Know what exists, grant access narrowly, check each action, and watch for the exception. The agents that breached Hugging Face escaped their evaluation environment, chained several vulnerabilities, and used credentials they found along the way to move through production systems at machine speed. If you're building something that closes the gap between what an agent can reach and what it should be allowed to do, we want to hear from you.
¹ Hugging Face, security incident disclosure, July 2026 (huggingface.co/blog/security-incident-july-2026), and "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident," July 27, 2026. The HDF5 read exposed pod credentials; the template injection gave code execution. huggingface.co/blog/agent-intrusion-technical-timeline
² OpenAI, "OpenAI and Hugging Face partner to address security incident during model evaluation," July 21, 2026 (updated through August 26), and "The Hugging Face incident and the road ahead," August 26, 2026. The models were GPT-5.6 Sol and an unreleased internal model, run in an evaluation based on the ExploitGym benchmark; the proxy was third-party software hosted internally by OpenAI. openai.com/index/hugging-face-model-evaluation-security-incident/ and openai.com/index/hugging-face-incident-and-the-road-ahead/
³ METR and Redwood Research, "Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident," August 26, 2026. metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
⁴ Cisco, Securing the Agentic Frontier: A Defense in Depth Strategy with Cisco SASE, solution brief, updated May 13, 2026. cisco.com/c/en/us/products/collateral/security/secure-access/securing-the-agentic-frontier.html
⁵ Cisco, Relentless Defense: A global report about cybersecurity in the AI era, 2026. Survey of 8,000 security professionals across 30 markets, conducted April and May 2026. cisco.com/c/m/en_us/solutions/security/cybersecurity-reports/relentless-defense-report.html
⁶ Peter Bailey, "Securing the Agentic Workforce: Cisco Announces Intent to Acquire Astrix Security," Cisco Blogs, May 4, 2026 (updated June 29, 2026). blogs.cisco.com/news/cisco-announces-intent-to-acquire-astrix-security
⁷ Peter Bailey and Ev Kontsevoy, "Identity Everywhere: Bringing Infrastructure Identity to Agentic IT," Cisco Blogs, August 25, 2026. blogs.cisco.com/security/identity-everywhere-bringing-infrastructure-identity-to-agentic-it. Teleport product description: goteleport.com/platform/unified-identity-layer/
⁸ Arcade, "Deploy Agents Even Your Security Team Will Love," arcade.dev/get-started/authorization/, and "Production-Ready AI Agents for Financial Services," arcade.dev/blog/build-ai-agents-for-fintech-payments/. Product descriptions are Arcade's own.
⁹ "AI Agents Need New Security: Cisco Announces Intent to Acquire WideField Security," Cisco Blogs (updated July 31, 2026). blogs.cisco.com/news/cisco-announces-intent-to-acquire-widefield-security