Agents aren't magic—they're a security and cognitive liability

Tracy Bannon warns that AI agents in browsers and local apps create massive attack surfaces. Logging into a site and authorizing an agent gives it keys to the kingdom. Most users don't map data flows or understand cross-boundary risks. Hyperscalers may be safer within their ecosystems, but crossing tenants or platforms is dangerous.

The real bottleneck isn't model quality but workflow design and governance. Agents can beget agents, leading to proliferation without control. Teams face cognitive overload as they shift from authors to reviewers of AI-generated content. Volume of output outstrips human capacity to validate, creating a vicious cycle.

Testing agents is fundamentally different from testing deterministic code. Repeatability and auditability are unclear when token counts and tunnel vision affect outcomes. Specifications exist but testing methodology lags. The software development lifecycle, optimized for humans, doesn't fit agent-driven generation.

Skepticism is warranted around claims of groundbreaking security tools like Anthropic's Mythos. Earlier models found similar vulnerabilities. The hype exceeded practical capability. The real concern is societal: outsourcing thinking, loss of retention, and dependence on models controlled by a few.