Descriptions:
Idan Goldman, co-founder and CEO of Anchor, walks through the full stack of challenges that emerge when AI agents need to authenticate as humans on real websites—a problem his company has been solving in production with partners including Cloudflare, 1Password, fingerprint.com, and Coinbase. Presented at AI Dev 26 in San Francisco, the talk uses the metaphor of an iceberg to explain why identity for browser agents looks simple at the surface but hides enormous complexity underneath.
Goldman structures the talk as a progressive layering of problems. The naive approach—passing credentials as plaintext in a prompt—exposes secrets to external LLMs. The next step, obfuscating credentials at runtime, is better but still insufficient. Getting past web application firewalls and CAPTCHAs required Anchor to build a full verification stack in partnership with Cloudflare and fingerprint.com, allowing Anchor agents to be recognized as legitimate bots rather than blocked by protective systems. The network fingerprinting problem—where AWS IP addresses are automatically blocklisted by most websites—led Anchor to build its own enterprise VPN in partnership with AT&T, providing stable, high-quality IP addresses that pass SSO and Okta consistency checks.
Goldman frames the entire talk around the concept of “complexity creep,” illustrated by Anchor’s own mistake of building a custom billing system that ballooned far beyond its initial scope. His advice for teams adding browser agent capabilities is to treat identity as a deep, evolving problem rather than a checkbox—and to avoid owning layers of the stack that are not core to their product. The talk is essential viewing for teams building computer-use agents that need to operate reliably inside real enterprise authentication environments.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Deep Dive
