A Sandbox Is Not a Data Permission Model: Governing Enterprise AI Agents
A sandbox bounds agent code, not entitlements. How to add scoped agent identities, row and column controls, audit logs and revocation.
An agent can run inside a hardened sandbox, never touch the host kernel, and still read every row its service credential can reach. That gap is where most enterprise agent programs stall: the team has shipped sandboxed agent execution and believes least-privilege data access came with it. It did not. Runtime isolation constrains what agent code can do to the machine it runs on: which processes it reaches, which network paths it opens, whether it can escape to the host. Data authorization decides which records, columns and objects a given identity may read or write once a query reaches the data platform. A sandbox bounds the first and is silent on the second.
Perplexity's own announcement of SPACE, the sandbox platform behind its Computer product, illustrates the boundary precisely. The company open-sourced SPACE and reported that its published security evaluation found no escape from the virtual machine to the host in 108 attempts, though some testers found ways around network restrictions (Perplexity, "Open-sourcing SPACE", 2026-10-15, vendor-reported). That is a meaningful result for runtime isolation. The same post says nothing about identity, authorization or data permissions.
The tension is structural, not a vendor oversight. Two control planes answer two different questions, and most enterprises have staffed only one. The sections below separate them, describe what a vendor sandbox does and does not guarantee, and set out a control set for scoped identities for AI agents, row-level and column-level access control for agents, agent audit logs and revocation, closing with an implementation checklist you can run against your own deployment.
Why sandboxed agent execution and least-privilege data access are two different control planes
Sandboxed agent execution and least-privilege data access are separate control planes because one governs the code's runtime environment and the other governs the identity's entitlements. Runtime isolation answers a narrow question: can this code reach the host, the network, or another tenant? Data authorization asks a different one: may this identity read this row, this column, this object? A deployment can pass the first test and fail the second, because a sandbox constrains code, not entitlements.
NIST's zero-trust architecture model separates the decision from the connection. A policy engine makes and logs the grant decision, a policy administrator executes it by commanding policy enforcement points, and the enforcement point enables, monitors, and terminates the connection (NIST SP 800-207, August 2020). That document describes a general enterprise architecture and never mentions AI agents; mapping its components onto agent deployments is an inference, not a claim the standard makes.
Control plane | Question it answers | NIST SP 800-207 component it resembles |
|---|---|---|
Runtime isolation | Can this code reach the host, the network or another tenant? | Policy enforcement point terminating the connection |
Data authorization | May this identity read this row, column or object? | Policy engine deciding, policy administrator executing |
OWASP's agent security cheat sheet puts authorization on the second plane explicitly: enforce it "in the execution component, outside the agent's context," and grant agents only the minimum tools their task requires (OWASP's agent security cheat sheet, retrieved October 2026). A valid sandbox paired with a broad service credential, an illustrative combination rather than an observed incident, passes the isolation test and fails the authorization one.
What Perplexity's SPACE sandbox actually guarantees, and what it leaves out
SPACE guarantees isolation of code execution and reports nothing about data permissions. The SPACE launch post states that no attempt escaped the virtual machine to the host in 108 tries, while some attempts found ways around network restrictions, since fixed (the SPACE launch post, 2026-10-15). The same post reports sandbox creation three to five times faster than the previous provider on the same production traffic, measured on Perplexity's own traffic with no published test harness and no third-party audit.
The sandbox boundary ends where the credential begins.
Independent coverage repeats the vendor's numbers rather than testing them. Semafor's report on the release gives the escape figure as 118 tries, attributes it to Perplexity, and adds that models broke out of other, non-open-sourced sandboxes (Semafor's report on the release, 2026-10-15). The 108-versus-118 discrepancy is unresolved in the coverage, and no independent verification of either figure exists.
Deployment guidance narrows the claim further. CoreWeave's deployment guide for SPACE describes an open-source Apache-2.0 platform that runs code in isolated Firecracker microVMs, with each sandbox a Kubernetes Pod and node services running privileged with host networking (CoreWeave's deployment guide for SPACE, read October 2026). The same guide states that the install path for existing clusters is a development and evaluation path, not a production deployment, and that it uses SPACE's test key broker, which is not designed to secure encryption keys. None of this speaks to which rows an agent may read.
How should an enterprise give an AI agent its own scoped identity?
Give the agent its own non-human identity with its own grants, and treat any session that borrows a person's login as a delegated case with separate controls. Snowflake's agent identity documentation splits the two modes: delegated agents act on behalf of a user, and the session belongs to the user while the platform marks it agent-active so policies and audit views can treat the traffic as agentic; autonomous agents act under their own identity and authorization through a SERVICE_AGENT user (Snowflake's agent identity documentation, retrieved October 2026). That user type has been generally available since July 2026 and supports workload identity federation, including SPIFFE and SPIRE (generally available since July 2026, 2026-07-23).
This is one vendor's implementation, not industry consensus, and scoped identities for AI agents remain the exception in practice. VentureBeat's June 2026 survey of enterprises found that only about a third give every agent its own scoped, managed identity, while most still let agents share credentials (VentureBeat's June 2026 survey of 107 enterprises, 2026-07-23). The wave is a single self-selected sample, the credential questions were multiple-select so the shares overlap, and the authors describe the results as directional.
A scoped identity is worth little if the credential never expires. SPIFFE workload identity issues SVIDs that are valid for a limited period, primarily to mitigate the likelihood of a key compromise, and the specification sets no universal lifetime (SPIFFE workload identity, living specification). Set the lifetime yourself and keep it short enough that a leaked token has a bounded window.
How do you start read-only and then add row-level and column-level controls for agents?
Start with a read-only scope that applies only while an agent is active, then layer row and column predicates on top of it. Snowflake's agent identity documentation describes the gate: "When a session is agent-active, the IS_AGENT_ACTIVATED context function returns TRUE," which you can test with SYS_CONTEXT('SNOWFLAKE$CURRENT','IS_AGENT_ACTIVATED')::BOOLEAN (Snowflake's agent identity documentation, retrieved October 2026). That predicate is the whole trick behind row-level and column-level access control for agents: the policy changes what is returned, not what the role is allowed to do.
The same check composes across control types. A masking policy can hide a sensitive column whenever an agent is active, and the identical expression works in row access, projection, aggregation and join policies (agent-aware masking policies, retrieved October 2026). Where a table carries both a row access policy and masking policies, the row access policy is evaluated first. This is one vendor's documented implementation, not an industry standard.
Pro Tip: Confirm the read-only scope holds in a test session before you add a single predicate. A ceiling you have not observed is a ceiling you do not have.
Without a scope set, the default is implied trust: the agent inherits the user's access. NIST's adversarial machine learning taxonomy (March 2025) recommends designing on the assumption that prompt injection is possible and differentiating permissions as a mitigation (NIST's adversarial machine learning taxonomy, March 2025). OWASP's agent security guidance recommends per-tool read-only versus write scoping and human approval for high-impact actions (OWASP's agent security cheat sheet, 2025), and its excessive agency risk names excessive permissions as a root cause (OWASP's excessive agency risk, 2025). No source prescribes numeric approval thresholds; define your own and write them down.
What should agent audit logs contain, and how do you revoke access quickly?
Agent audit logs and revocation both hang off the identity: a log that cannot name the agent cannot drive a clean revocation. Snowflake's implementation shows the two halves. The agent_type column in query history identifies the type of agent that directly invoked a query, with values for Cortex agents, lite agents and external agents, and NULL when no agent was involved; it rolled out over about a week from 27 July 2026 (the agent_type field in query history, 2026-07-27). In access history, the 2026_07 bundle removes invoker_identity and replaces it with agents_info, an ordered array of invoking agents from the nearest to the top-level agent, each carrying type, identifier and name (the agents_info column in access history, 2026_07). Dashboards built on the old column break when the bundle is enabled.
Revocation runs through the identity, not through key rotation. NIST's zero-trust architecture model treats it as an explicit capability: on denial, countermand or security alert, the policy administrator signals the enforcement point to shut the connection (NIST's zero-trust architecture model, August 2020). The NIST Generative AI Profile, voluntary and cross-sectoral, asks organizations to inventory third-party entities with access to organizational content and to cover access removal in deactivation plans (the NIST Generative AI Profile, July 2024).
The incident record shows why both halves matter. Hugging Face's July 2026 disclosure describes an intrusion driven end to end by an autonomous agent system that ran many thousands of actions across a swarm of short-lived sandboxes, harvested cloud and cluster credentials, and moved laterally into several internal clusters (Hugging Face's July 2026 disclosure, July 2026). That is one case, not a rate. The Agent Incident Registry, a cumulative and quality-filtered catalogue, holds 529 verified incidents as of 22 September 2026, of which 200 involved an agent doing real damage to a real party (the Agent Incident Registry, 2026-09-22).
Key Takeaway: If you cannot name the agent in the log, you cannot revoke it cleanly. Attribution and revocation are the same control, read in two directions.
Six controls to check against your own deployment:
Identity type. Every agent runs as a non-human identity with its own grants, not a borrowed login or a shared service account.
Session ceiling. A privilege ceiling applies whenever an agent is active, and you have verified it is not overridden for users with their own session policy.
Policy predicate. Row and column predicates key off the agent-active state, not off a separate copy of the data.
Attribution. Query and access logs carry the agent type and the agent chain, and no dashboard still reads the retired identity column.
Revocation. Disabling the agent identity ends its access, and you have tested that path rather than assuming it.
Approval thresholds. You have written down which actions require human approval, because no source supplies a number for you.
Two postures are defensible. Teams that started with runtime isolation have a hardened execution environment and an open question about entitlements. Teams that started with data authorization have scoped identities and an open question about what the agent's code can reach. Most enterprises need both, sequenced by which risk is live today.
Frequently Asked Questions
Is a sandbox enough to secure an enterprise AI agent?
No. A sandbox constrains what agent code can do to its runtime environment, and Perplexity's SPACE evaluation reports no virtual-machine-to-host escape in 108 attempts, which says nothing about which rows the agent's credential can read (the SPACE launch post, 2026-10-15). Data authorization is a separate control plane that decides what the agent's identity may access.
Can I use a sandbox and least-privilege data access together?
Yes, and they are designed to be layered. OWASP's agent security cheat sheet asks for authorization enforced outside the agent's context alongside sandboxing for arbitrary code execution (OWASP's agent security cheat sheet, retrieved October 2026). The sandbox bounds the code; the privilege ceiling and data policies bound the entitlements.
How do I revoke an agent's access quickly?
Revoke the identity, not the keys. NIST's zero-trust model has the policy administrator shut down the connection when an approval is countermanded, which in practice means disabling the agent's own identity and letting the enforcement point drop its sessions (NIST's zero-trust architecture model, August 2020). Keep credential lifetimes short so a leaked token expires on its own.
Is there a standard for agent data authorization yet?
Not a finished one. NIST's concept paper on agent identity, published as a draft in February 2026 with comments open through 2 April 2026, is exploring how agents should be identified, how delegation should be revoked and how logging should link agent actions to a non-human identity (NIST's concept paper on agent identity, February 2026). The IETF's on-behalf-of delegation draft, which illustrates a token carrying user, client and actor claims, was never an RFC (the on-behalf-of delegation draft, 2025). Treat current practice as recommendations drawn from vendor documentation and practitioner guidance, not compliance requirements.
Conclusion
A sandbox and a data permission model solve different problems, and the SPACE release makes the gap easy to see: a vendor can report zero host escapes in 108 attempts and say nothing about identity, authorization or data permissions in the same post. The evidence for the second control plane is thinner than the evidence for the first. One self-selected survey of 107 enterprises, one cumulative incident registry and one first-party post-mortem describe the risk; no source measures how much a privilege ceiling reduces it.
What the sources do support is a sequence. Give the agent its own identity, attach a read-only ceiling that applies only when an agent is active, layer row and column predicates on the same agent-active check, make sure the logs name the agent and the object, and test revocation through the identity rather than through key rotation. NIST's zero-trust model supplies the architecture for that sequence, OWASP's cheat sheet supplies the enforcement placement, and the platform documentation supplies the field names. None of them supplies your approval thresholds, which is why the checklist asks you to write them down.
The next step is a review, not a purchase: take the six controls above, walk your production agent deployment against each one, and note where the answer is "we assume so" rather than "we tested it." Those gaps are the data authorization work your sandbox never covered.
About the publisher. Singdata is a Singapore-headquartered company that builds data agents on an AI-ready unified lakehouse platform, including workspace-based access control and audit of data operations. It has a commercial interest in this topic, so weigh its perspective accordingly. The guidance above does not depend on its products.
find out more
singdata.com/trending/sandbox-not-data-permission-model
singdata.com/trending/patch-public-before-hole-commit-to-advisory-gap