SELECT * FROM users is an attack when it reaches a support chatbot and a normal request when it reaches a coding agent.

One prompt injection model can't tell the difference, which is why we trained one for each environment. launch.arcjet.com/Q4Ew6I4

One prompt injection model can't protect every AI agent. A SQL query is an attack on a support chatbot and a normal request to a coding agent. Today we're launching Arcjet's custom AI security models, each trained for the environment its agent works in. launch.arcjet.com/Q4Ew6I4

Arcjet Himalia 1 is our general prompt injection model, and it is live now for every Arcjet customer. Arcjet Thebe 1 for coding agents and Arcjet Carme 1 for autonomous security agents are in early access.

The models are compact classifiers fine-tuned from Qwen3-0.6B, so they return a verdict in milliseconds at a fraction of the cost of asking a frontier model. Himalia 1 scored F1 results of 0.964 and 0.971 on two public prompt injection datasets.

We're at the AI Security Summit in San Francisco today. Stop by the Arcjet booth to see real-time threat analysis and session behavior analysis protecting coding agents. Your developers will keep choosing their own tools, and the policy should follow them.

Arcjet logs every coding agent call with the agent, the user, the prompt or tool call, the policies that ran and the decision, which is more context than an AI gateway can see. Send it to Splunk, Datadog, SentinelOne or Panther. hubs.li/Q04zy-pl0

Developers use Claude Code, Codex, Cursor, Muse Code, Copilot and more, often within a managed enterprise account, but not always. Arcjet detects shadow AI through coding agent hooks, OpenTelemetry and provider APIs, with an optional macOS agent. hubs.li/Q04zy-r50

Great fun building the @arcjet app using the native macOS network extensions to detect shadow AI for today's launch of runtime security for coding agents.

We use a decision model with the full session context to detect credential exfiltration, destructive commands, behavior ctivity, all in real time.

Reach out for a demo or sign up to try it today!

All the details: arcjet.com/blog/introduci…

Today we're announcing Arcjet Runtime Security for Coding Agents. Security teams can now discover all the coding agents used across every developer endpoint and apply policies to secure those activities, whatever coding agent is being used. hubs.li/Q04zy-nY0

Read a ticket. Query the customer database. Send an email.

Each action is permitted. The sequence is an exfiltration path. Per-call checking cannot catch this no matter how good the individual checks are, because the earlier steps are out of scope by the time it matters.

Observe, enforce, audit. Thirty seconds on what we announced last week and why enforcement and evidence belong together.

Your agent is not waiting for someone to type into a chat box. It is set off by an inbox, a text message or a code commit, and it reads and writes to your database on the way through.

That is why we built enforcement that runs at the moment it acts.

hubs.ly/Q04xJdvT0

Agentic AI can now take real actions - its mistakes range from clumsy to harmful to malicious.

@arcjet's new Agent Runtime Security gives engineering teams enforcement as soon as an agent acts, with sequence-aware checks so risky patterns get caught.

It's a year since I moved to NYC, so we're on Times Square for today's launch of @arcjet's new agent runtime security:

Three new capabilities, available now:

1. Observe: discover all your agents, wherever they're running, with no code changes and nothing to deploy

2. Enforce: nd after every action. Apply deterministic policies to tools, APIs, database calls, and other inputs/outputs. Detect prompt injection, wrap tools in safety guardrails, limit API access, and protect PII

3. Audit: collect the evidence for security reviews and compliance audits

Sign up for a free trial or reach out to schedule a demo!

When AI was just a chatbot, the worst thing it could do was be wrong. That is no longer where the risk sits. Today we are announcing Agent Runtime Security.

Audit. Every decision is recorded with its full context, so a security review or compliance audit becomes a query rather than a reconstruction.

Enforcement is also sequence aware, so a decision late in a workflow accounts for what happened in the earlier steps instead of judging each call on its own. Full write-up in the reply. launch.arcjet.com/gcX0rBr