AI Agents Get Cheaper Faster Than We Can Secure Them

The second gap concerned not access rights but confidentiality. Anthropic, OpenAI, and Google encrypt reasoning traces—the model’s step-by-step reasoning, which the API does not show directly to users so competitors cannot copy the model’s thought process through distillation. Researchers showed that the encrypted traces of all three companies could be reproduced in significantly weaker models and obtained in unencrypted form.

The vulnerability has been patched, but for months it made it possible to extract market leaders’ intellectual property through an ordinary public API.

For a company connecting an LLM to internal systems through MCP or its own gateway, both incidents point to the same lesson: the price and convenience of AI tools are improving far faster than the maturity of the processes around them. With GPT-5.6 costing 80% less, running an agent on real tasks is becoming cheaper than hiring a junior employee.

But an inexpensive agent with broad access becomes an inexpensive incident with broad access when boundaries are not explicitly defined:

• which systems the model can see
• which commands it can execute
• what happens to its output on the way to a competitor
• to another company’s server

Isolating agent code is a concrete engineering task: CPU and memory limits, network access denied by default, and a filesystem whitelist—exactly what smolvm provides for untrusted code. The same principle applies to model access to internal systems at the request level: an LLM & Security Gateway sits between the agent and product systems, logging and restricting calls, while MCP integrations define access explicitly, tool by tool, rather than using a single token with unrestricted permissions.

The integrator does this work before releasing the agent to production—securing the perimeter in advance instead of patching the pilot after an incident.

In 2026, AI-agent infrastructure is becoming cheaper and simpler before our eyes—this is a fact, not a forecast. Every story about greater convenience is paired with one about a breached perimeter: another company’s system compromised during a test, or someone else’s intellectual property extracted through an API. A company that cuts costs on a security gate to launch an agent faster pays for that saving precisely when the cost of a mistake becomes public.