Compact Model Detects Prompt Injection in AI Agents
Arcjet's classifier-based approach is available now, with two additional models in early access and no published benchmark yet.
Updated on Nov. 6, 2026 in Artificial Intelligence
Live Poll
Do you trust AI assistants to safely handle your data and accounts?
Arcjet has released Arcjet Himalia 1, a prompt injection detection model now available and powering detection across its platform. Arcjet Thebe 1 and Arcjet Carme 1 are available in early access, so the three models are at different stages of availability.
Why it matters
Prompt injection attacks can embed malicious instructions in web pages, emails, code repositories and other content that AI agents encounter, potentially redirecting agent behavior toward exposing credentials, modifying code or triggering unauthorized actions. Arcjet says it built its models after finding limitations in existing open-source prompt injection detectors, reflecting a wider push to secure agents as they take on more autonomous work.
Arcjet's initial prompt injection models are fine-tuned from Qwen 3-0.6B, a 600 million parameter open-weight model. The source reports no accuracy, false-positive or latency numbers, so the models' detection performance relative to existing open-source detectors remains unmeasured in the available material.
The players
Arcjet
A security company whose platform now runs prompt injection detection for AI agents, integrating into coding agents and custom agents in code.
The details
Arcjet uses compact, fine-tuned classifiers rather than large general-purpose language models to evaluate incoming requests. The models are trained on manually reviewed public datasets plus proprietary data, and use a classification architecture that outputs a security verdict without generating text. Arcjet combines this specialized behavior classification with session-level analysis, so suspicious patterns can be judged across a series of interactions rather than a single request.
Timeline
Arcjet detailed its new family of AI security models in early November 2026.
The Tech Race
Arcjet's approach of compact classifiers over large general-purpose models places it among efforts to secure AI agents with narrow, purpose-built detectors. Its claim to improve on existing open-source prompt injection detectors is the competitive point to track against independent benchmarks.
Developers building coding agents and custom agents can integrate Arcjet's detection into their code, with Himalia 1 available now and Thebe 1 and Carme 1 in early access. Pricing and rollout timelines for the early-access models were not stated in the source.
The takeaway
Watch whether Arcjet publishes independent benchmark results comparing Himalia 1 with existing open-source prompt injection detectors. Movement of Thebe 1 and Carme 1 out of early access will show how far the compact-classifier approach extends beyond the first model.
Further reading
For broader context on securing AI systems, explore our coverage of Artificial Intelligence.
Source note: This article includes information reported by Computer Weekly.
Live Poll
Do you trust AI assistants to safely handle your data and accounts?
