gbrindisi

oh I love the thought that there is a future where I get my models burned in hardware same as I used to get games in cartridges

experimenting with a bunch of things:

- reimplemented claude code auto mode for pi

- prompt injection detections refactored from llm judges to classifier judge

- as above various first pass detections for generated and untrusted files, such as moderation and malware detection

- ranking search results in web_search tools to reduce context overload

I found that a lot of what i was using a one shot haiku llm call for can now be done with jev or another decision model.

Not just sandboxing but overall security engineering practices on both sides

why is that?

Shameless plug: in case someone finds this requiring too many tokens, we shared the recipe on how we built our own in house audit skill so that it can easily be replicated and tuned to different environments https://www.synthesia.io/post/automating-code-security-revie...

Claude, write a skill to create design docs as per this blog post /s

In all seriousness a big chunk of my work is reviewing docs and 100% of the design docs I read these days are generated and there is a lot of slop. I have to parse them with AI to make sense of them, contributing to the problem. I wish people would follow a sensible set of rules like this one.

Did anyone have any solution to this dread? Or cracked the code on how to write docs that have other agents as their primary audience?

it kinda does unless he is straightforward about his financial incentives

I used to accept only people I worked with/met in person, but the feed is crap anyway so why bother. I now accept pretty much everyone and prune the feed.

I agree. They have quality data to build an effective AI pentest product that is good enough, and they already have a good offering to bundle that into and satisfy enterprise demand.

Up and coming AI pentest companies need to have an exceptional product to get a chance to stand on their own and penetrate the enterprise market, otherwise their best scenario is an acquisition to get bundled into an established platform.

I also notice the growing trend to have EM carry individual contributor duties, I thought it was mostly a consequence of using coding agents but perhaps it's not: the EM figure as we know might just be a consequence of the golden zirp times (do you remember the endless technical EM vs non-technical EM debates?)

It's not just AI, these are the consequences of affiliate marketing. Just look at the crap that is YouTube nowadays.

I like the pattern of making a dedicated cli/harness and just build a skill to teach coding agents to use it.

At $work we built a thorough workflow to do security reviews, which is a pure skill to simplify adoption https://www.synthesia.io/post/automating-code-security-revie...

But the user experience is tricky because if we aim for very low false positives the run time for this kind of workflows is too long, it's then hard to justify blocking PRs.

I protest the modern web by trying to consume all content via RSS.

The feed reader shall be my main window to the world, and I am sorry that it's not obvious to content creators that I read them so I often send an email on the note of "I enjoyed this article you wrote, thanks".

I write a small blog myself and I see the other side of it, but I just gave up on SEO, metrics, etc. I want to be the change I'd like see in the world: I publish full content RSS, I remove all analytics, make the website as lean as I can, put out my contact data and my only success metric is # of interactions I get with occasional readers.

This is pretty much a spec driven workflow.

I do similar, but my favorite step is the first: /rubberduck to discuss the problem with the agent, who is instructed by the command to help me frame and validate it. Hands down the most impactful piece of my workflow, because it helps me achieve the right clarity and I can use it also for non coding tasks.

After which is the usual: write PRDs, specs, tasks and then build and then verify the output.

I started with one the spec frameworks and eventually simplify everything to the bone.

I do feel it’s working great but someday I fear a lot of this might still be too much productivity theater.

are agents/ still relevant after we got skills? I am genuinely confused on why I would need custom system prompts for specific agents, what should I use them for?

thanks for raising the alarm and sharing this, very insightful

(also beautifully presented!)

1. I dont have hard metrics at hand but with the latest Sonnet I'd say we reach consensus around 80% of the time, with Opus is almost always but we are not using it due to cost

2. The difference I see in agent behavior when they don't reach consensus is usually either

- when one of them didn't explore enough and lack context

- and/or when their risk assessment is off

The latest happen often, in other workflows based on agents we are now giving clear instruction on how to assess risk and where to draw a line to consider something a true positive.

3. validation is on Sonnet, we don't use persona based prompts but all the 3 validators get's the same task and context. The agent orchestrating them will take their output and make the final decision. We use an internal fork of the claude code github action for now.

I am doing something similar: I use openspec to create context and a sequential task list that I feed to ralph loops, so that i’m involved for the planning and the verification step but completely hands off the wheel during code generation.

I like openspec, it lets you tune the workflow to your liking and doesn’t get in the way.

I started with all the standard spec flow and as I got more confident and opinionated I simplified it to my liking.

I think the point of any spec driven framework is that you want to eventually own the workflow yourself, so that you can constraint code generation on your own terms.

fifteen years ago I use to do mobile pentests for banks and when we could not find anything significant for the reports we could’ve always count on “lack of rooting detection” and pin the risk on some vague mobile banking malware threat pushed by marketing. I am sorry I contributed to this nonsense.

100% security theater, and here we are.

ah I also did my own sandbox and at least twice the agent inside tried really hard to go around the firewall, so I ended up intercepting calls to `connect` to return a message that says "Connection refused by the sandbox, don't try to bypass".

Code here: https://github.com/gbrindisi/agentbox

the most annoying thing with Google Workspace is that you need super admin privilege to properly audit the environment programmatically, I believe because of the cloud-identity api.

I noticed that too and it’s kinda scary. Soon we will have the opposite of canceling, where the target will be deepfaked to say everything and its opposite to nullify their signal to noise ratio.

The crowdstrike incident taught us that no one is going to review any dependency whatsoever.