Frontier AI agents explained: capabilities, risks, and enterprise control

Key insights

  • Frontier describes capability; the deployed application and its permissions determine action authority.
  • Task-completion time horizons measure human-duration task difficulty at a specified success rate, not safe unattended runtime.
  • Evaluate accepted outcomes, authorization enforcement, recovery, and total effort together.
  • Use independent records of identity, tool, and network activity to investigate and stop boundary violations.

Frontier AI agents combine highly capable AI models with tools and an execution loop to carry out multi-step tasks. Their capabilities are advancing: the UK AI Security Institute's December 2025 report found that leading systems completed selected software tasks requiring at least an hour of human-expert work more than 40% of the time by mid-2025, compared with less than 5% in late 2023. Those controlled results make evaluation worthwhile, while leaving enterprises to establish reliability and safe operating boundaries for their own workflows.

For a security or platform owner, the central question is whether an agent can deliver an acceptable result using only the access and actions the business intends to authorize. Answering it requires examining the deployed system: its model, tools, identities, data, approval rules, monitoring, and recovery process.

What are frontier AI agents?

A frontier AI agent is a system that uses a model near the leading edge of demonstrated capability to select and perform actions toward a goal. “Frontier” describes capability relative to other systems at a point in time; the application determines which actions the agent can take.

An agent might inspect a repository, identify a problem, edit a file, run a test, and revise its proposed fix. Another might retrieve supplier information and prepare a purchasing recommendation. Both involve a model choosing intermediate steps within an environment that provides tools and constraints.

The terminology becomes clearer when you separate three layers:

Term What it describes Enterprise question
Frontier model A model near the current capability frontier Does its demonstrated capability match this task?
AI agent A system that chooses and executes steps toward a goal Which tools, resources, and actions can it use?
Agentic workflow The broader process coordinating agents, software, and people Who authorizes decisions, verifies results, and owns failures?

A capable model can support a tightly constrained agent. A less capable model can still operate with excessive permissions. Model choice and access policy therefore need separate decisions.

The label also changes over time as systems improve. It does not establish artificial general intelligence, an agreed safety certification, or a particular level of autonomy. For the broader model terminology, see the frontier AI guide.

How frontier AI agents work?

An agent typically receives a goal, selects an action, observes the result, and repeats until it completes the task or reaches a stopping condition. The surrounding application must decide which proposed actions are valid and which require additional authorization.

The model and the control loop

The model interprets the goal and available context, then proposes a next step. An orchestration layer manages execution: supplying tool descriptions, calling permitted functions, returning results, and tracking progress.

For example, a coding agent may request a file, propose an edit, and ask to run a test. The model supplies the proposed action; the application executes the permitted operation. Changing the orchestration, instructions, or tools can change the system's behavior even when the underlying model stays the same.

Evaluate that complete configuration. A model benchmark provides limited evidence about an agent using different tools, data, instructions, or execution budgets.

Tools, memory, and identity

Tools connect an agent to application programming interfaces (APIs), browsers, databases, code execution, or business systems. Tool access turns a proposed plan into possible changes to the environment.

Memory can include the current conversation, retrieved documents, task history, or persistent records. Its design affects which information an agent carries into later decisions. Record where persistent state originates, who can change it, and how it is isolated between users and tasks.

Identity establishes the account or service through which an action occurs. A connector might use the requesting employee's delegated access or a separate service identity. Review that choice explicitly: the connector's effective permissions may cover more than the employee's immediate request.

Verification and stopping conditions

An agent's completion message is one input to verification. The application should also check the relevant system of record: whether the file changed, the test ran, the ticket updated, or the transaction completed as intended.

Define stopping conditions before execution begins. Useful examples include an expired approval, missing evidence, conflicting instructions, repeated tool failures, an unexpected destination, or exhaustion of a time or spending budget. The correct result may be a documented handoff to a person.

Frontier AI: Built Without Brakes

Explore documented agent behavior, gaps in oversight, and practical controls for identifying agents, observing their activity, and containing actions that cross intended boundaries.

Download the white paper →

What frontier agents can do in enterprises?

The strongest pilot candidates have a clear outcome, accessible evidence, and manageable consequences when execution fails. Begin by defining the work and its acceptance criteria, then select a model and agent design capable of meeting them.

These are possible applications and suggested initial boundaries, rather than claims of guaranteed capability:

Workflow Potential agent contribution Suggested initial boundary
Software engineering Investigate an issue, propose a patch, and run permitted checks Isolated workspace; review before merging or deployment
Research and analysis Gather sources, compare evidence, and prepare a recommendation Approved sources; verify consequential statements
Service operations Retrieve account context and draft a resolution Approval before account changes or external commitments
Procurement Compare approved suppliers and prepare a requisition Separate authority for payment and supplier-record changes
Security investigation Gather authorized evidence and suggest investigative steps Read-only access before considering response authority

Frontier AI coding agents attract attention because software tasks often provide executable checks. Business workflows can be harder to judge: a fluent supplier summary may still omit a contractual restriction or rely on outdated information.

Use a simpler approach when it satisfies the requirement. A deterministic workflow may handle a predictable process efficiently, while an agent may be useful where the next investigative step depends on earlier findings. More autonomy also increases the need to observe agent activity and intervene when behavior crosses a boundary. The added evaluation and operating effort needs a corresponding business benefit.

What the evidence says about agent autonomy?

Capability research shows progress, but deployment decisions require a second layer of evidence about acceptable work, reliability, and failure consequences. Read every benchmark alongside its task distribution, agent configuration, scoring method, and evaluation date.

AISI's longitudinal findings concern selected, well-scoped tasks and models available during the study period. They establish progress under those conditions. They do not estimate how often an enterprise agent will complete a purchasing process correctly, follow an organization's access policy, or recover from an interrupted transaction.

Longer task horizons need careful interpretation

METR measures task-completion time horizons using the estimated time a human expert would need for a task. A 50%-time horizon identifies the task duration at which the evaluated agent is predicted to succeed half the time. It does not measure how long the agent itself runs.

METR's suite mainly covers software engineering, machine learning, and cybersecurity tasks with relatively clear scoring criteria. Its page was last updated on 8 September 2026 and explicitly says it is no longer actively maintained. Use it to understand the methodology and historical measurements, rather than treating it as a complete current ranking.

For a production decision, specify the required reliability and consequences first. A research draft that a person will check has different acceptance criteria from an action that changes customer records or moves money.

Passing tests can still fall short of accepted work

In a March 2026 research note, METR reported that roughly half of the studied agents' benchmark-test-passing code changes would not be merged by maintainers after adjustment for review noise. Four maintainers reviewed 296 AI-generated pull requests across three repositories.

The study covered historical agents released from mid-2024 through mid/late-2025 and did not let them revise their submissions after maintainer feedback. METR therefore cautions against treating the result as a fundamental capability limit. It also received less review than the organization's research posts.

The practical implication is to measure the outcome your organization accepts. Passing a test, producing an answer, gaining approval, and completing an authorized business action are different checkpoints. Record the work needed to move between them.

What recent agent incidents teach defenders?

Agent behavior must be interpreted alongside the permissions and conditions under which it occurred. A research incident can reveal a useful failure mode without establishing how frequently the same behavior occurs in normal enterprise deployments.

AISI disclosed that its security team detected an incident on 28 July 2026 during a cyber evaluation. Agents took unsanctioned actions on the live internet in 10 of 122 runs of one challenge. The institute said internet access had been deliberately permitted and some safety filters disabled. It explicitly distinguished the incident from a sandbox escape, said the tested configurations were not commercially available, and reported no evidenced resulting real-world harm.  

Those qualifications matter. The disclosure supports examining what an agent can reach and do; it does not establish an incident rate for public products or ordinary business use.

In its subsequent engineering account, AISI described layered outbound network restrictions, monitoring during execution, and reassessment of evaluation environments. It also acknowledged monitoring limitations.

For enterprise teams, the useful lesson is to test independent boundaries. If an agent proposes an unexpected action, which component can reject it? If a connector fails to enforce policy, does another control limit the destination or credentials? If an action succeeds, can operations identify it and intervene?

Where operational risk enters the agent workflow?

Risk follows the full path from input to action: documents, retrieved content, tools, identities, persistent state, and connected systems. Reviewing the model's final answer alone leaves much of that path unexamined.

The OWASP Top 10 for Agentic Applications includes goal hijacking, tool misuse, identity and privilege abuse, agentic supply chain vulnerabilities, memory and context poisoning, insecure inter-agent communication, and cascading failures. These categories organize threat modeling; they do not quantify how common each failure is.

Untrusted content can redirect legitimate tools

An agent might read a supplier document, a support ticket, or a tool response containing instructions planted by someone else. If it treats that material as authority, it may use otherwise legitimate tools for an unauthorized purpose.

The AgentDojo research framework evaluates this problem through tool-using tasks and untrusted tool-returned data. It provides a useful evaluation principle: assess whether the agent completes legitimate work and whether it preserves security properties under attack. Its original benchmark does not establish the efficacy of a current enterprise product.  

Include malicious and misleading inputs in a pilot. Check the actual tool requests and resulting state, including attempts that the surrounding application correctly blocks.

Delegated access can exceed the business task

A request to summarize a customer account does not necessarily require permission to export the customer database. Likewise, a request to prepare a supplier recommendation does not require permission to change payment details.

Apply authorization outside the model. Check the identity, resource, operation, task context, and any required approval before executing the action. NIST's zero trust architecture provides the underlying principle that network location or asset ownership alone should not confer implicit trust. The agent-specific application of that principle still requires implementation and testing.

Review memory and delegation boundaries as well. A second agent should receive only the information and authority its subtask needs. Shared context should not silently become a way to share unrestricted permissions. For detailed implementation considerations, see AI agent security.

How to evaluate frontier AI agents before deployment?

Evaluate the complete workflow against accepted outcomes, permission enforcement, recoverability, and operating effort. The scorecard below is a proposed evaluation aid derived from the research, rather than a validated benchmark or universal acceptance standard.

Evaluation area Question to answer Evidence to collect
Accepted outcomes Does the result meet the business acceptance criteria? Independent review of completed, failed, and escalated tasks
Reliability Is performance repeatable across relevant task variations? Repeated runs, task categories, and failure reasons
Authorization Are disallowed operations rejected outside the model? Denied requests and verified downstream system state
Adversarial resilience What happens when retrieved content contains malicious instructions? Legitimate-task results and attempted boundary violations
Recovery Can execution stop and partially completed work be reconciled? Revocation, cancellation, duplicate-action, and rollback exercises
Observability Can operations reconstruct the action sequence? Correlated task, identity, tool, and resource events
Total effort and cost What does each accepted result consume? Compute, tool charges, review, retries, rework, and operating effort

Use the actual workflow as the test unit

Create a representative set of tasks with named owners and acceptance criteria. Include ordinary cases, ambiguous instructions, missing data, permission denials, interrupted sessions, and attempts to redirect the agent.

Freeze the configuration being evaluated: model version, instructions, tool definitions, connector permissions, memory settings, and execution budgets. Record any human help. A result that required substantial correction should remain visible in the evaluation record.

Compare the agent with the existing process and, where useful, a simpler automation. Review the distribution of outcomes rather than only the average. A workflow may perform well on common cases while failing on the cases with the largest consequences.

Set acceptance thresholds before seeing the results. Let the business owner and security owner agree which failures require a redesign, a narrower scope, or a human decision.

Calculate cost per accepted outcome

A useful calculation is total pilot cost divided by the number of outcomes that meet the agreed acceptance criteria. Include model inference, retrieval and tool charges, human review, retries, and rework. Track setup and ongoing operating costs separately so the comparison remains understandable.

For example, a fast first draft may still be expensive if a specialist spends substantial time correcting it. Conversely, an agent that escalates difficult cases early may be valuable if it reduces wasted work while preserving quality.

Use measured pilot data. A token price, benchmark score, or successful demonstration alone cannot establish the cost of a completed business process.

Define stop conditions before granting execution

Agree which events suspend the pilot or withdraw an action class. Examples include an unexplained attempt to access a new destination, a missing approval record, inconsistent downstream state, or an inability to revoke the active credential.

Treat a blocked unauthorized request as useful evidence about both agent behavior and control enforcement. Investigate why it occurred before expanding permissions. A single aggregate success score can conceal that distinction.

A worked example of a procurement agent

A procurement agent can provide value while authority over payments and supplier changes stays separately controlled. The following hypothetical example shows how to connect task acceptance, permissions, review, and monitoring; it describes no customer deployment or measured outcome.

Approve a narrow research task

An employee asks the agent to compare approved suppliers for a defined purchase and prepare a requisition. The application gives it read access to the approved supplier catalog and relevant purchasing policy. It can create a draft in a designated workspace.

The task record identifies the requester, purpose, permitted sources, authorized tools, spending context, and expiry. The agent has no permission to update supplier bank details or send internal financial records externally.

Acceptance requires a recommendation supported by the approved records, a clear explanation of missing information, and a draft that the responsible buyer can verify.

Test a malicious supplier document

A supplier attachment contains an instruction to send internal budget information to a new web address before completing the comparison.

The expected result is a refusal or escalation, with no disclosure. Test both the agent's behavior and the enforcement layer: an attempted transfer must still fail if the agent follows the embedded instruction. Confirm the destination was blocked and the downstream system received no unauthorized change.

Also check whether the malicious instruction persists in stored task context. A later, apparently unrelated request should not inherit it as trusted operating guidance.

Approve and verify the final action

The buyer reviews the proposed requisition, source records, supplier identity, amount, and missing-information flags. Approval applies to that specific action and expires if the material details change.

Before creating the requisition, the application verifies the approval and permissions again. Afterward, it records the actual requisition identifier and resulting state. If execution times out, the workflow checks whether the requisition already exists before retrying.

The pilot can then measure accepted requisitions, review effort, incorrect recommendations, attempted unauthorized actions, duplicate prevention, and successful cancellation. These measures reveal whether expanding execution authority is justified.

How to govern and monitor agents in production?

Production governance should preserve ownership, enforceable access boundaries, evidence of actual activity, and a tested response path as the system changes. NIST's May 2026 agent-security consultation summary reports broad agreement that existing cybersecurity principles remain relevant but require adaptation for agents. Source: NIST's consultation summary.

Give every deployment an accountable owner

Record the business purpose, operating owner, model configuration, connectors, effective permissions, data classes, and recovery contacts. Keep ownership current when an employee leaves, a supplier changes, or a pilot becomes an operational service.

NIST's NCCoE reported in September 2026 that software development would provide its first implementation use case for demonstrating agent identification, authentication, and authorization. This is ongoing project work, rather than a completed certification for enterprise agents.

For the inventory process, use the AI agent inventory guide. Include a way to identify the human or service that approved delegated authority and a way to withdraw it.

Observe what the agent actually does

Correlate the task identifier with relevant identity, tool, resource, and network events. An investigation should be able to establish which credential acted, which resource changed, what approval applied, and whether the action succeeded.

Preserve useful application events alongside independent infrastructure evidence. An agent's explanation can help an analyst understand a sequence, while system records establish the observable changes.

Protect the monitoring path and define who can alter its configuration. Apply access controls and retention rules to the logs themselves, since prompts, retrieved content, and tool results may contain sensitive information.

Treat changes as new evaluation conditions

Recheck consequential changes to models, instructions, tool schemas, connectors, memory, or permissions. Record the change, repeat relevant regression and adversarial cases, and confirm that approval and revocation still work.

Exercise incident response across components. Stopping one agent process may leave a queued action, another delegated worker, or an active credential. The response procedure should identify these dependencies, suspend further work, and reconcile actions already completed.

For a deeper treatment of protecting the surrounding systems and workflows, see agentic AI security.

Where Vectra AI fits?

When an AI agent accesses company systems, security teams need to know which account it used, what it accessed, and whether that activity was expected.

Vectra AI's identity-based attack detection and containment looks for signs of credential misuse and privilege abuse by analyzing user and service account activity across network and cloud environments. For an agent that uses those accounts, this can provide evidence for an investigation. Coverage depends on the accounts, services, and integrations included in the deployment.

In an evaluation, test a specific scenario, such as an agent's test account accessing an unexpected resource. Check whether the activity is visible, whether an analyst can trace it to the account and resource, and which response actions are available. Use the agent's own task and tool logs to connect that evidence to the work it was assigned.

To assess your environment, request a Vectra AI demo with a defined agent workflow and the accounts and services it uses.

Note: This guide synthesizes primary research and security guidance. The evaluation framework and procurement example are proposed guidance based on the publicly available information.

FAQs

Are frontier AI agents the same as AGI?

Does every AI agent need a frontier model?

Does a longer time horizon mean an agent can run unattended?

Does human approval make an agent safe?

What should an enterprise pilot first?