New researchSeptember 2026

    Secure the agent, not just the model.

    AI agents turn model outputs into persistent actions. This research maps where enterprise agent stacks break, the controls that close those gaps, and measured security results for Pokee-Isaac 28B v0.1 (preview).

    Christopher Wu, Zhaorun Chen, Michael Cai, Bo Li, Zheqing Zhu

    Read the full whitepaperOpen Trust Center

    Benchmark

    DecodingTrust-Agent

    Same 12 domains and judge

    8.3%

    Indirect injection success

    Lower is safer

    30.5%

    Direct attack success

    Lower is safer

    85.6%

    Benign task success

    Higher is better

    ~6,200

    Tasks per model

    Same 12 domains and judge

    Model comparison

    Same 12 domains · same judge

    DecodingTrust-Agent results for Pokee-Isaac 28B v0.1 (preview) and five comparison models
    ModelIndirect injectionLower is saferDirect attackLower is saferBenign task successHigher is betterTasks per modelApproximate
    Pokee-Isaac 28B v0.1 (preview)8.3%30.5%85.6%~6,200
    Claude Haiku 4.536.7%37.7%70.5%~6,200
    Nemotron 3 Super39.9%78.3%61.0%~6,200
    GPT-5.6 Luna46.0%52.3%84.9%~6,200
    Qwen3.5-122B46.1%59.8%77.9%~6,200
    Gemini 3.5 Flash-Lite49.0%82.9%82.5%~6,200

    The attack surface

    Traditional app security sees events. Agent security must see trajectories.

    A benign request can encounter a poisoned email, accept a malicious tool return, combine two individually safe permissions, and make an irreversible change. The risk only becomes visible when the complete sequence is preserved.

    1. 1

      User objective

      The authorized task and intended outcome.

    2. 2

      Untrusted content

      External data enters the working context.

    3. 3

      Planner decision

      The agent selects a path and its next action.

    4. 4

      Tool composition

      Permissions combine across tools and steps.

    5. 5

      Persistent action

      A change lands in an enterprise system.

    Model and planner

    Direct harmful requests and conflicting instructions can redirect planning before a tool is ever called.

    Tools and skills

    Tool schemas, MCP metadata, reusable skills, parameters, and returned content all become part of the decision surface.

    Retrieval and memory

    Poisoned documents or persistent memories can change decisions across later steps, tasks, and sessions.

    External environments

    Webpages, email, chat, files, and third-party systems can carry instructions the user never authored or saw.

    Serving infrastructure

    Identity, batching, caches, budgets, telemetry, and model APIs create security boundaries below the application layer.

    Hardened architecture

    Eight controls. Two owners. One defensible system.

    Three controls sit below the API and require the agent platform. Three belong to the customer’s security layer. Two are joint. Neither party can secure the deployment alone.

    SM1Security layer

    Policy gateway

    Model-independent input classification, direct-harm refusal, parameter rejection, and identity propagation.

    SM2Joint

    Injection firewall

    Preserve provenance on every return, classify injected instructions, and detect drift from the user’s task.

    SM3Security layer

    Capability broker

    Authorize composed capabilities, not just individual endpoints, with egress allowlists and trust tiers.

    SM4Platform

    Output guard

    Guard the stream above the sampler for degenerate output, secret leakage, and PII egress.

    SM5Platform

    Isolation

    Use per-caller budgets, tenant-tagged buffers and cache keys, plus application-level wall-clock bounds.

    SM6Security layer

    Audit plane

    Redact at write time and enforce retention, residency, and access controls on the investigation record.

    SM7Joint

    Supply chain

    Sign skill manifests, track corpus provenance, and quarantine untrusted memory before it becomes persistent.

    SM8Platform

    Configuration attestation

    Compare source defaults with deployed configuration and fail builds when security controls silently diverge.

    Measured results

    Injection resistance without giving up the work.

    Across the same 12 DecodingTrust-Agent domains and with the same judge, Pokee-Isaac 28B v0.1 (preview) beats all five comparison models on every reported measure: indirect injection success, direct attack success, and benign task success.

    Read security relative to capability

    A low attack-success rate is not meaningful if an agent rarely completes benign work. The report keeps BSR beside every ASR result so refusal, failure, and genuine resistance are not conflated.

    • Results are from one run; the paper applies a ±3-point noise band.
    • Pokee-Isaac 28B v0.1 (preview) runs in its own agent scaffold; the other five models run in the stock OpenAI SDK runner.
    • All six models use the same 12 domains and the same judge, with approximately 6,200 tasks each.
    36-page whitepaper

    Read the architecture, methodology, per-domain results, and caveats.

    Download PDF

    Cite this work

    BibTeX

    @techreport{wu2026enterprise,
      title       = {Enterprise Agent and Model Security: Threat Landscape, Hardened Architecture, and Measured Results},
      author      = {Wu, Christopher and Chen, Zhaorun and Cai, Michael and Li, Bo and Zhu, Zheqing},
      institution = {Pokee AI},
      type        = {Whitepaper},
      year        = {2026},
      month       = sep,
      url         = {https://pokee.ai/research/pokee-ai-security-whitepaper.pdf}
    }

    Review Pokee’s operational security posture.

    The research explains the architecture. The Trust Center contains current compliance documentation and assurance materials.

    Visit Trust CenterSecurity overview
    Pokee Logo

    Pokee AI

    Enterprise AI agents, deployed in your infrastructure.

    Solutions

    FinanceHealth CareE-commerceLawEducationManufacturingTechnology

    Company

    CareersSecurityContact Us

    Follow Us

    TwitterLinkedInRedditDiscord

    Resources

    API Documentation

    Legal

    Terms of ServicePrivacy PolicyAccessibilitySystem Status

    © 2026 Pokee AI. All rights reserved.

    Terms & ConditionsPrivacy Policy