AI Governance Institute
← News
Research2026-09-02

Canva's CISO: Default Trust in AI Agents Is an Enterprise Control Failure

Source

Why I Dont Trust Your AI Agent | Kane Narraway, Canva

AI Security Podcast

Via AI Security Podcast

What happened

In an episode of the AI Security Podcast titled "Why I Don't Trust Your AI Agent," Kane Narraway, Canva's CISO, made the case that AI agents should not be assumed trustworthy simply because they come from established vendors. The conversation centered on how rapid vendor proliferation in the agent space has outpaced enterprise evaluation capabilities, leaving organizations approving agent deployments without the adversarial scrutiny those deployments warrant. Narraway's framing positions agent distrust as a principled default rather than a one-time procurement checklist item. The discussion is part of a growing practitioner consensus that agentic systems require a fundamentally different security posture than conventional software — one where the question is not whether a vendor is reputable but whether the agent's behavior, permissions, and tool access have been independently validated. This episode lands in a period of sustained attention to agentic control failures, including findings that half of enterprises cannot trust their AI agents' decisions and that two-thirds lack agent governance policies as network-layer controls begin to emerge.

Why it matters

  • ·Vendor reputation does not substitute for behavioral validation in agentic deployments. Compliance teams that rely on a vendor's published safety documentation as the primary intake control are operating with an assurance gap that regulators and auditors are increasingly likely to scrutinize, particularly as agent governance is becoming binding across multiple jurisdictions.
  • ·Agent tool use and plugin connections expand the third-party risk surface well beyond the primary vendor relationship. Each tool server, API integration, or plugin an agent can access is a potential trust-chain failure point, and enterprises without explicit scope and permission controls are exposed to lateral movement risks that conventional vendor due diligence processes were not designed to detect.
  • ·A default-distrust posture for AI agents requires operational changes that many governance programs have not yet codified. Pre-deployment adversarial testing, explicit permission manifests, and behavioral boundary documentation need to become standard intake conditions rather than optional maturity-level enhancements.

Stay ahead of stories like this

Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Governance controls affected

What to do now

  • ☐Audit every active third-party AI agent deployment to confirm that a behavioral evaluation — not just vendor documentation review — was completed before go-live.
  • ☐Add a mandatory adversarial testing gate to your agent intake workflow that requires evidence of scope-boundary testing before any agent with tool-use or external API access is approved.
  • ☐Review agent permission manifests for all deployed agents and revoke any permissions that are not explicitly required for the documented use case.
  • ☐Update vendor due diligence questionnaires to include agent-specific questions covering tool access controls, permission escalation prevention, and sandbox isolation capabilities.
  • ☐Brief your AI governance committee on the default-distrust posture and confirm that your current agent approval process reflects this standard rather than relying on vendor safety attestations alone.

What to watch next

Compliance teams should monitor whether practitioner consensus around default-distrust for AI agents begins to harden into formal regulatory expectations, particularly as national guidance bodies issue more specific pre-deployment evaluation requirements. The trajectory of agentic governance rulemaking, including binding standards emerging from the agent governance landscape and CISA's agentic identity and approval standards, suggests this posture will move from best practice to baseline expectation within the next regulatory cycle. Teams should also track whether insurance underwriters begin conditioning AI-related coverage on documented adversarial evaluation of agent deployments, which would create a financial incentive parallel to the regulatory one.

Related Coverage

Corporate Policy2026-11-19

Five Major US Banks Disclaim AI Agent Liability, Including After Hacking

Citibank, JPMorgan, Bank of America, PNC, and Citizens have each updated their personal account terms of service. The updates hold customers solely responsible for losses caused by third-party AI agents granted account access. Citibank's policy goes further, placing responsibility on customers even when the AI tool is hacked and their credentials are stolen by attackers. The updates affect personal account holders who authorize AI tools to access their accounts, not enterprise or business banking accounts.

Corporate Policy2026-11-22

Treasury Nominee Leaves Agentic AI Bank Liability Unresolved at Senate Hearing

A Treasury Department nominee declined to define who bears responsibility when a third-party AI agent takes unauthorized actions in a consumer bank account. The nomination hearing before the Senate Banking Committee on November 20, 2026 produced no regulatory commitment beyond a pledge to coordinate with existing bank regulators. The Bank Policy Institute separately stated that agentic AI risk management is not sufficiently developed to support account access, naming prompt injection as a concrete unresolved threat.

Corporate Policy2026-11-21

Apple's macOS Agent Plans Rewrite Endpoint Access Control Assumptions

Apple is reportedly developing a first-party AI agent for macOS capable of browsing the web, running commands, writing code, and operating across apps. Simultaneously, Apple plans to introduce friction in macOS 27.3 that discourages users from granting third-party agents full system access. Neither the first-party agent nor the macOS 27.3 controls are yet in force. Bloomberg reports the in-house agent is still in early development. A launch is possible no sooner than the end of 2027.