WHY APPLICATION-LEVEL PERMISSIONS WON'T SECURE AI AGENTS WITH OKTA'S & NANOCO | DISRUPT 2026 - OKTA INC

Financial Performance:

• No specific financial metrics (revenue, growth, profit margins, cash flow, deferred revenue, or debt) were disclosed in the transcript as NanoClaw is in the early startup phase (founded ~Jan 2024) and Okta's specific quarterly earnings figures were not the focus of the discussion.
• Okta confirmed it participated as a strategic investor in NanoClaw's seed round.

Guidance and Future Outlook:

• Immediate Trend: Agentic software is shifting from individual employee utility to "AI teams" or swarms of agents collaborating, with projected scaling from dozens to hundreds or thousands of agents per team within the next 3–6 months.
• Market Timing: Founders are advised to anticipate that the security and technical challenges of today may be obsolete within 4–6 months due to rapid model advancements; strategies must be built for the future state, not current capabilities.
• Product Roadmap: NanoClaw is preparing to launch a free consumer version following its enterprise deployment, focusing on intent-based authorization and agent-to-agent interaction security.
• Long-term Goal: Establishing a future where humans direct AI agents rather than working for them, requiring robust governance, least-privilege infrastructure, and clear accountability structures.

Business Segments and Product Lines:

• NanoClaw: Focuses on "agentic software" security, utilizing a "containerization" and "locking" approach to isolate agents from the core environment while enforcing policies externally; features include a kill switch for sessions, tools, and entire infrastructure, and intent-based approval workflows.
• Okta: Evolving beyond identity verification for humans to managing identity for AI agents, treating agents as "insider threats" with equivalent onboarding and access governance needs; offering "XAA" (Extended Access Authentication) standards and partnerships via the "Blueprint Alliance."
• New Capabilities: Both companies are addressing "agent-to-commerce" interactions and the protection of personal agents that employees might bring into the corporate network.

Market and Competitive Landscape:

• Market Sentiment: Enterprises feel "vast majority... behind" in AI adoption but are rushing to integrate agents without establishing necessary security primitives (data access mapping, toolkits, roles).
• Competition/Standards: The landscape is defined by the race to implement "least privilege" and deterministic access; Okta's XAA standard is cited as a key enabler for connecting agents securely.
• Threat Evolution: The threat landscape is accelerating, with risks becoming "deterministic" and "real" every six months; the emergence of low-cost, high-capability open-weight models increases the frequency of potential insider threats and exploits.

Risks and Challenges:

• Insider Threat: The primary risk is agents behaving as malicious insiders by exploiting access to data lakes, BI systems, or spreadsheets if not properly isolated or monitored.
• Infrastructure Debt: Organizations face "time bombs" from years of unpatched CVEs (Common Vulnerabilities and Exposures); agents can exploit these historical vulnerabilities to hop between systems like "Swiss cheese."
• Isolation Failures: Risks include container escapes, VM escapes, and the inability to implement effective "kill switches" on unpatched infrastructure or agents running on personal laptops.
• Compliance Risks: Specific examples of SOX compliance compromise were noted due to unauthorized data exports to public AI models.

Management Commentary and Tone:

• Tone: Urgent, cautious, yet optimistic about the potential of AI; management emphasizes that security cannot be an afterthought and that "don't trust" agents means "trust but verify" with external enforcement.
• Confidence: Leaders express strong confidence in the necessity of shifting from ad-hoc security to deterministic, intent-based authorization and full infrastructure isolation.
• Philosophy: A consensus on "human-centric AI" where humans define intent and AI executes, contrasting with the fear that agents will eventually make decisions humans must work for.

Analyst Questions and Answers:

• Q: Why is "don't trust AI agents" important?
  • A: It is not about suspending use but recognizing that autonomy requires external, verifiable enforcement outside the agent's environment to prevent misbehavior or policy violations.
• Q: How do Okta and NanoClaw differ in approach?
  • A: Okta takes a "broader aperture" on enterprise identity and access management (IAM) for agents, while NanoClaw focuses on isolating the agent in a "box" and enforcing policies on tools/actions externally; both aim for least privilege but operate at different layers of the stack.
• Q: What is the biggest security mistake companies make?
  • A: Rushing to value extraction without mapping data access, toolkits, and roles upfront, leading to uncontrolled exposure; or locking down access so completely that employees resort to shadow IT (exporting data to public models) which is less secure.
• Q: Advice for founders?
  • A: Build for the market's reality 4–6 months from now, not today; skip the "small startup to enterprise" ladder and aim for the enterprise security stack immediately; ensure N-1 patching programs are active.

Other Key Points:

• Kill Switches: Essential at multiple layers: session termination, tool access revocation, and infrastructure-wide shutdown (inference access cuts); currently, most enterprises are far from implementing meaningful kill switches due to legacy infrastructure.
• Procurement Reality: Large enterprise security reviews and procurement processes take 3–6 months; this friction cannot be skipped, and AI enablement leaders often lack deep enterprise experience to navigate it.
• Best Practice: Implement an "N-1 patch program" immediately to address the "swarm" of vulnerabilities agents can exploit; assume VM/container escapes will happen and build least-privilege infrastructure accordingly.
• Future Architecture: The shift involves moving from individual agent interactions to "agent teams" (swarms), necessitating new security models for inter-agent communication and collective decision-making.