Augustus LLM Vulnerability Scanner Augustus is a brand-new, open-source vulnerability scanner made to protect Large Language Models (LLMs) from a constantly changing array of hostile attacks. Praetorian created Augustus, a single-binary solution that can launch more than 210 different adversarial attacks against 28 LLM providers, with the goal of bridging the gap between academic research tools and production-grade security testing. Security teams have struggled with tooling that is frequently research-oriented, slow, or challenging to integrate into continuous integration/continuous deployment (CI/CD) pipelines as businesses scramble to incorporate Generative AI into their products.
Although they depend on intricate Python environments and numerous dependencies, current tools such as NVIDIA's Garak have established the benchmark for thorough testing. By being compiled as a single, portable Go binary, Augustus overcomes these operational bottlenecks.
By doing away with the need for virtual environments, pip installs, or particular interpreter versions, this architecture removes the "dependency hell" that is frequently connected to Python-based security tools. The tool is substantially faster and more resource-efficient than its predecessors because it uses Go's built-in concurrency primitives, or goroutines, to carry out massively parallel scanning. In their release announcement, Praetorian said, "We needed something built for the way our operators work: a fast, portable binary that fits into existing penetration testing workflows."
210+ Modes of Attack Augustus is fundamentally an attack engine that makes the process of "red teaming" AI models automated. It comes with a library of more than 210 vulnerability probes covering 47 types of attacks, such as: Jailbreaks: Advanced prompts (like DAN, AIM, and "Grandma" exploits) made to get past security measures.
Prompt Injection: Methods for overriding system instructions, such as Base64, ROT13, and Morse code encoding bypasses. Data Extraction: Reconstruction of training data, API key disclosure, and PII leakage tests. Examples of adversarial tactics include logic bombs and gradient-based attacks that aim to perplex model reasoning.
Augustus' "Buff" system, which enables security testers to dynamically apply transformations to any probe, is one of its best features. To test whether a model's safety guardrails withstand obfuscated inputs, testers can chain several "buffs," such as paraphrasing a prompt, translating it into a low-resource language (like Zulu or Scots Gaelic), or encoding it in poetic formats. This feature is essential for identifying "fragile" safety filters, which might stop a typical attack but miss the same attack when it is slightly modified.
Built for the current security stack, Augustus comes pre-configured with support for 28 LLM providers, including local inference engines like Ollama and well-known platforms like OpenAI, Anthropic, Azure, AWS Bedrock, and Google Vertex AI. Because of this extensive support, teams can use the same tools to test anything from locally installed Llama 3 instances to cloud-hosted GPT-4 models. In order to prevent scan failures during extensive evaluations, the tool's architecture prioritizes production reliability and includes built-in rate limiting, retry logic, and timeout handling.
It is simple to ingest vulnerability data into vulnerability management platforms or SIEMs because results can be exported in a variety of formats, such as JSON, JSONL for streaming logs, and HTML for stakeholder reporting.
The LLM fingerprinting tool Julius was released first, and Augustus is the second in Praetorian's "12 Caesars" open-source series. The Apache 2.0 license makes it instantly available. For daily cybersecurity updates, developers and security experts can download the most recent build or release from source via X, LinkedIn, and GitHub.
To have your stories featured, get in touch with us.
