tp tate@programs field note
tate@programs ~/notes/agent-security-evidence-2026 may 2026

field note / agent security

Agent demos need evidence, not vibes.

The 2026 agent-security question is no longer just whether the model can answer. It is whether a reviewer can inspect what the agent was allowed to read, write, call, send, spend, deny, and escalate.

surface
agent demos
risk
tools / data / spend
output
evidence pack
updated
May 2026

practical read

The trust layer has to be visible.

Agent teams are moving from chat demos into workflows that touch private files, SaaS APIs, MCP servers, payment flows, browsers, databases, and other agents. That creates a different launch bar. A useful demo is not enough if the reviewer cannot see the boundaries.

The public TechEx / lablab enterprise AI hackathon tracks point at the same gap: agent security, AI governance, Gemini-based agents, and deployable enterprise workflows. The interesting part is not the phrase "agent security." It is the evidence buyers and judges can inspect.

What an evidence pack should show

  • What prompt-injection and exfiltration drills were run.
  • Which policy rule matched the risky step.
  • Whether the action was allowed, denied, rate-limited, quarantined, or sent to human review.
  • Which tool scope was active: read, write, network, filesystem, payment, or destructive action.
  • What the agent claimed it intended to do versus what the policy layer detected.
  • Which audit event proves the decision after the demo is over.

A2A makes identity part of security

The Agent2Agent project frames A2A around agent discovery, capability declaration, task collaboration, and secure interoperability. That means an agent-security review should not stop at prompt text. It should also ask: who is this agent, what skills does it declare, what provider owns it, and what authentication is expected before another agent delegates work to it?

Gemini belongs in the test loop, not in the leak path

Gemini and AI Studio are useful for generating adversarial drill cases, summarizing evidence, and testing long-context agent workflows. The operational mistake is putting a model key where the browser bundle, a screenshot, or a public repo can expose it. A safe demo uses AI Studio directly or keeps API keys behind a server-side route.

Policy without audit is hard to sell

A policy rule that blocks a bad action is useful during the demo. An audit event that explains the block is useful after the demo. The second part is what makes the difference for regulated workflows, enterprise buyers, and judges looking for measurable risk reduction.

The small workflow that helps most

  1. Load the project or paste the architecture notes.
  2. Run prompt-injection, exfiltration, tool-boundary, and human-review checks.
  3. Generate the fix queue.
  4. Export a policy starter, A2A Agent Card, audit schema, and Gemini drill prompt.
  5. Keep the evidence pack with the submission or buyer handoff.

Why I built the drill kit

Agent Security Drill Kit is a browser-only version of that workflow. It is intentionally local, because early-stage teams should be able to inspect a project without uploading source code to a third-party scanner. The current version exports a review pack that can be used for hackathon submissions, launch reviews, and product demos that need a concrete trust story.

open https://tateprograms.com/agent-security-drill.html
load project
export evidence pack
patch top risks

source trail

Signals this page tracks.

techex

Agent Security & AI Governance

The May 2026 TechEx / lablab page centers enterprise agent security on guardrails, observability, access control, audit, and red-team tooling.

open source

lobster

Policy actions

Lobster Trap positions prompt inspection around policy actions, metadata, audit logs, and deployable enforcement for agent workflows.

open source

a2a

Agent identity

The A2A project describes agent discovery, capability negotiation, task collaboration, and security expectations between independent agents.

open source

gemini

Model testing

Google's Gemini API and AI Studio docs are the integration path for teams using Gemini in agent workflows and test generation.

open source