Adversarial Minds

Agentic AI you can defend — to your security team, your board, a courtroom. 25 years securing the stack. Teaching CAIOs the layer down.

[Adversarial Minds @adversarialmnds]

@listelia Interested

[Adversarial Minds @adversarialmnds]

With all of the new threats associated with the latest frontier models, agent swarms and quantum replay attacks, should we be re-thinking data at-rest?

How can we build context systems that rely upon just-in-time and other techniques to avoid aggregation of data that is difficult to protect?

[Adversarial Minds @adversarialmnds]

I can see some truly useful things that Jev can do, especially in the area of speeding up judgment and making it more accurate. But I can't help thinking that folks are just forcing the new shiny thing into any old use case.

Where does it truly shine?

[Adversarial Minds @adversarialmnds]

Agents improvise around ambiguous instructions. If a bad option is the easy one, they will take it.

[Adversarial Minds @adversarialmnds]

If the agent can read the secrets file, the keys are already in the chat.

[Adversarial Minds @adversarialmnds]

@DanielMiessler Is AI improving the lives of dolphins?

[Adversarial Minds @adversarialmnds]

Do you know whether your Hermes Agent's work is successful? Tokens, tool calls, and response time are easy to count. The harder question is whether they produced the right result.

[Adversarial Minds @adversarialmnds]

Can you tell whether a change actually improved your Hermes Agent? My first test treated three wins in five runs as verification. A useless change can pass that standard half the time.

[Adversarial Minds @adversarialmnds]

Let's face it. Bryce has to grow up. He's still thinking he's all that and he's not doing the introspection work needed to truly understand the game.

His demeanor on the sideline looked like a pouting little kid, not someone who has humility and is determined to understand what he truly needs to do to grow into his potential.

[Adversarial Minds @adversarialmnds]

Can a faster Hermes Agent be a worse Hermes Agent? Two of my five faster test runs failed because the agent skipped context it needed. Speed improved. The answers did not.

[Adversarial Minds @adversarialmnds]

I've actually built a plugin to attempt to assess this in Hermes Agent. It evaluates on four planes: Outcomes, Cost, Latency and Hygiene and it builds statistics based upon the model it runs on. Results from my own testing so far indicate that Astra is a little bit better than GLM 5.2, but there are still agentic fundamentals to worry about. Deterministic controls are still needed.

[Adversarial Minds @adversarialmnds]

@ZackKorman @ZackKorman I've seen your latest posting on rogue AI agents. But an agent is going to act on its owner's intent. What's your definition of rogue?

[Adversarial Minds @adversarialmnds]

@matviy It doesn't have to suck, it just needs to be optimized to your use. I've spent time analyzing what makes my Hermes Agent tick. Watch this space for my forthcoming article and dashboard plugins.

[Adversarial Minds @adversarialmnds]

Before you optimize your Hermes Agent, can you trust the baseline? My first measurement reported a 39.9-second median because one 14-hour build session counted as a single turn. The math was fine. The population was wrong.

[Adversarial Minds @adversarialmnds]

The agent will retry a broken call forever if you let it.

[Adversarial Minds @adversarialmnds]

Does your Hermes Agent repeat work it has already done? Mine was making roughly 144 duplicate tool calls per week. Every duplicate added time and cost without improving the answer.

[Adversarial Minds @adversarialmnds]

Do you know what makes your Hermes Agent slow? Thinking, tool calls, and retrieved context all add time and tokens. I've been measuring which ones help, which ones waste resources, and how to tell the difference.

[Adversarial Minds @adversarialmnds]

@freddiexpott Folder

[Adversarial Minds @adversarialmnds]

Does your Hermes Agent ever take several minutes to answer a single prompt? Mine was taking six to eight. The answers were usually good, but I wanted to know what happened during those minutes and which steps actually needed to happen.

[Adversarial Minds @adversarialmnds]

Extra tool calls make AI agents slower and more fragile. Every unnecessary step costs time and creates another place for the task to fail. Agent efficiency is part of reliability.