Arcjet AI security models - A family of custom-trained security models for coding agents
by•
Arcjet is launching a family of custom AI security models designed for analyzing AI agent activity: Arcjet Himalia 1, a general purpose prompt injection model, live today for every SDK call that uses prompt injection detection, and specialist versions designed for situations that general models handle badly, Arcjet Thebe 1 for coding agents, and Arcjet Carme 1, for autonomous security agents.
Replies
@davidmytton put it simply: "You can't detect prompt injections with a general-purpose model."
First launched here 2 months ago and loved by the community (Top Product of the Week, featured in the newsletter), @Arcjet keeps the momentum by announcing a family of custom-trained security models designed for analyzing AI agent activity:
Arcjet Himalia 1, a general-purpose prompt injection model (live now)
Arcjet Thebe 1, for coding agents (in early preview)
Arcjet Carme 1, for autonomous security agents (in early preview)
The models are compact classifiers fine-tuned from @Qwen 3-0.6B, so they return a verdict in milliseconds at a fraction of the cost of asking a frontier model.
More details about the benchmark
Get started at arcjet.com
Repost on X
What should @Arcjet build/improve/fix from your perspective?
Hey team, a huge milestone today! Arcjet Security Models came to change the AI game and I'm truly sure many founders will take the most of it. Keep it up!!!
@german_merlo1 Thanks! Try it out and let me know how you get on.
@german_merlo1 thanks for the support!! you rock
@tessak22 you're the 🐐
@fmerian thanks friend!
Thanks @tessak22 !
what happens when the injection gets the agent to write into CLAUDE.md and it only fires next session, does session analysis carry anything across sessions
@bbueno Arcjet's models are the first line of defense for prompt injection. When they detect a potential injection, a secondary decision model runs with a larger context that examines the whole session, including activities before and which action the agent is going to do next. This allows us to detect more than just basic prompt injection. Arcjet also supports ongoing session analysis so every action is assessed individually and in the context of the broader session. This allows us to detect agent behavior change over time - we can block risky actions like credential exfiltration or risky code execution even if they come later in a session.
@davidmytton makes sense, thanks David
Arcjet an essential company in the space right now. Excited to try out these models! How did you post-train them out of interest?
@samsmallmind Most of the work was in curating high quality datasets because many of the open examples are ambiguously licensed and poorly labeled. Our technical writeup is at https://arcjet.com/blog/introducing-arcjets-custom-ai-security-models
looks super interesting and relevant to me as a solo founder and for those who uses coding agents - good luck with the launch!
Thanks @withstephen - let me know if you'd like to give it a go so we can help protect your business.
congrats on the launch. to your question, i would want a plain log of why a request was blocked, so a false positive on a real signup is easy to debug )
@agshin_rajabov The Arcjet SDK returns the decision with the reason so you can see it in the context of the prompt that triggered it. We can also optionally retain the prompt on our platform so you can see the results of historic evaluations. We have customizable retention policies, including the option for zero data retention, so it's up to you what gets stored.
seeing the reason next to the prompt is what i needed, and zero retention as an option makes it much easier to try )
yo arcjet, you're one of my favorite devtools
congrats on the launch and bringing more goodness for agents security! 🔥
The Thebe 1 angle is the one that lands for me. I run coding agents against my own repos all day, and the stuff they read — docs, dependencies, random files — is exactly the kind of content that looks suspicious out of context. A general prompt-injection check would drive me nuts with false blocks, so a specialist model plus that second decision step looking at the whole session before blocking a risky action feels like the right trade-off. One thing I'm curious about: my agents mostly run locally, so does every check round-trip to Arcjet, or can the small classifier run close to the agent?
@jcbuilds we run the models in the cloud on our platform. We're regularly updating them with new data and manage the scaling and inference for you. Analysis happens at the edge though, and we're deployed on Cloudflare so have over 300 locations worldwide.
i'll share this with my builder friends