Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

INFERA AI.SafeCode maps eight scanners into an attack map and throws attack prompts at chatbots

A security report lists a hundred vulnerabilities, but it is unclear how far an attacker can reach from outside. INFERA AI.SafeCode combines data from eight scanners into a map that runs from a public endpoint through the code to a dangerous operation. The AI Red Team module sends attack prompts to a chatbot and calculates the attack success rate (ASR). So fix the findings on reachable paths first, and block releases when ASR grows.

AI-processed from Habr AI; edited by Hamidun News
INFERA AI.SafeCode maps eight scanners into an attack map and throws attack prompts at chatbots
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Five metricsASR, citation, grounded, hallucination rate, and ASR drift
Key metric #2The baseline is the previous version's behavior: even within the norm, rising ASR is a signal
Key metric #3The threshold is set and breached, so the build fails with a quality gate error

Picture a security report: over a hundred vulnerabilities, each with a CWE number, a file, a line, and a severity level. Where do you start? A vulnerability in an abandoned module and an identical one in a public endpoint look the same in the table, even though their risk differs.

The developers of the Russian platform INFERA AI.SafeCode described on Habr how they tackle this problem. The first approach: combine data from eight scanners into a single attack surface map, that is, everything an attacker can reach from outside.

The second approach: AI Red Team, which automatically tests a running chatbot or assistant with a stream of attack prompts, where the main metric is ASR (attack success rate), showing the share of successful attacks.

From a list of findings to a reachability map

The eight scanners are SAST (source code analysis), SCA (checking open-source dependencies), secrets detection, DAST (testing a running application), Pentest (an AI agent that imitates manual penetration testing), Code Fuzzing (searching for crashes and edge cases), API Fuzzing (REST, GraphQL, gRPC, and WebSocket), and an MLSecOps scan that identifies the components of an AI system and looks for defects in them. The authors themselves acknowledge that eight scanners in one product is not an advantage in itself. The value lies in linking the results together. A code finding is tied to an endpoint, an endpoint is tied to the external surface, and the status of a check is tied to a specific run and its date.

The most useful cut is the chain "input → path → sink". A sink here is a dangerous operation: an SQL query, command execution, unsafe deserialization, or exposure of sensitive data. Suppose you have a hundred findings. If there is a path from a public API to a vulnerable handler, that finding gets handled first. If there is no path, it drops lower in the queue. The call graph also shows risk concentration, for example several public endpoints sharing the same unsafe handler.

Blind spots are singled out separately: what was not checked. An empty report does not mean the system is secure if DAST was not run after routes changed, SCA is looking at a stale snapshot of dependencies, or API fuzzing does not know about new methods. That is why a coverage metric should answer not only "how much was scanned" but also "which important paths remain unchecked."

How an LLM application is tested

A regular web service has a request, a response, and expected behavior. An LLM application's output depends on the model, the system prompt, the tools, the retrieval data, and even how the question is phrased. A vulnerability shows up not as a crashed process but as incorrect behavior. The assistant reveals its system instructions. A protective rule is bypassed by rephrasing or a multi-step dialogue. The model calls a tool with dangerous parameters or answers confidently without verification. Classic scanners do not catch this.

AI Red Team sends the chatbot a corpus of attack prompts. This is a versioned set of scenarios, not manual input into the interface. In parallel, a quality corpus can be run: questions with reference answers. It is needed to check not only whether the model held its ground, but also whether it has become worse at answering on point.

The platform calculates five indicators:

  • ASR: the share of successful attacks;
  • citation rate: the share of answers with a required source reference;
  • grounded rate: the share of answers grounded in permitted data;
  • hallucination rate: the share of answers containing hallucinations;
  • ASR drift: the change in attack success relative to the baseline.

The baseline is the recorded behavior of the previous version. Its purpose is not to declare the system "safe forever" but to catch regressions. If ASR rises after a change of model, prompt, or retrieval index, that is a risk signal, even when the absolute value still falls within policy.

How to roll it out without stopping the pipeline

If thresholds are not set, a run remains just a measurement. Breaching at least one set threshold makes the run return a quality gate error (a quality check at the build exit), and the release can be halted. The authors advise moving in stages. First, observation on a test stand. Then a soft policy, where metrics are visible in CI but builds are not blocked. Hard blocks are enabled only for critical attack classes. After that, the corpus is regularly expanded with real incidents and results of manual checks.

An "ASR equals zero" threshold sounds strict, but for a broad corpus with ambiguous scenarios it is unrealistic. It is more practical to split attacks into classes and block releases on those that are critical for your business. Agree in advance on what counts as a successful attack. For one application, it is enough not to leak a secret; for another, you need to confirm that the model did not call a tool or alter a record.

What to keep in mind

The material was written by the platform's own developers. It contains no independent measurements, no figures on false positives, and no cost data for runs. A reachability graph does not replace manual analysis of business logic.

The practical takeaway is useful even without this specific product. Keep attack and evaluation corpora alongside the code. Tie results to the commit and the model version. Repeat Red Team testing after any change to the model, prompt, index, or tool set. Prioritize findings by reachability. For teams shipping LLM features, this is a way to turn security from a one-off "approval before release" into a repeatable test of every version.

Story

Key Questions Answered

What are the key details on From a list of findings to a reachability map?

The eight scanners are SAST (source code analysis), SCA (checking open-source dependencies), secrets detection, DAST (testing a running application), Pentest (an AI agent that imitates manual penetration testing), Code Fuzzing (searching for crashes and edge cases), API Fuzzing (REST, GraphQL, gRPC, and WebSocket), and an MLSecOps scan that identifies the components of an AI system and looks for defects in them. The authors themselves acknowledge that eight scanners in one product is not an advantage in itself.

How an LLM application is tested?

A regular web service has a request, a response, and expected behavior. An LLM application's output depends on the model, the system prompt, the tools, the retrieval data, and even how the question is phrased. A vulnerability shows up not as a crashed process but as incorrect behavior. The assistant reveals its system instructions. A protective rule is bypassed by rephrasing or a multi-step dialogue. The model calls a tool with dangerous parameters or answers confidently without verification.

How to roll it out without stopping the pipeline?

If thresholds are not set, a run remains just a measurement. Breaching at least one set threshold makes the run return a quality gate error (a quality check at the build exit), and the release can be halted. The authors advise moving in stages. First, observation on a test stand. Then a soft policy, where metrics are visible in CI but builds are not blocked. Hard blocks are enabled only for critical attack classes. After that, the corpus is regularly expanded with real incidents and results of manual checks.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…