INFERA AI.SafeCode builds an attack surface map and hits LLMs before release

⚡ The gist in 5 seconds

• INFERA AI.SafeCode assembles a project's attack surface from SAST, SCA, secret scanning, DAST, and reconnaissance of the running application.
• The map shows which API entry point can actually reach a dangerous sink: SQL injection, command injection, insecure deserialization, sensitive data leaks.
• A separate AI Red Team mode attacks an already running LLM application and tracks regressions between versions.

🔍 What happened

INFERA Security published a breakdown of two modes of its AI.SafeCode product on its Habr blog — an 8-minute read. At the time of publication, the article had 4.5K views and a rating of 25.3 with 11 channel subscribers. This is a vendor text about its own product, not independent research, and should be read with that caveat in mind.

Here's the mechanics. AppSec used to live in checkpoints: static code analysis, dependency checks, manual pentesting before release, occasionally dynamic scanning. The report came out as a list: CWE, file, line, severity, recommendation. AI.SafeCode merges the results of SAST, SCA, secret scanning, DAST, and reconnaissance of the running application into a single graph, adding a call graph and scanner coverage data. The team then looks at three views: a risk summary (how many nodes, entry points, and sinks, how many high and critical, which data is stale), a node inventory — a filterable list of files, symbols, endpoints, external addresses, and technologies — and the main map: API entry points connected by paths to vulnerable operations. For LLM applications, a second layer has been added — attacking the live AI interface: the model's resilience to unwanted prompts, response quality, and regressions between versions.

🧠 Our take

The logic is simple and, honestly, long overdue. A vulnerability in an experimental module that can't be reached from outside and the same vulnerability in a public endpoint are two different risks, yet the scanner assigns them the same severity. Our assessment: the main product here isn't "yet another scanner," but reachability-based prioritization. That's the part usually done by hand, taking eight hours if you're lucky.

What this changes for teams. The conversation with the business shifts from "we have 340 findings" to "three public endpoints lead to a critical sink, here they are." That's an argument that can be verified and challenged. In our experience, AppSec is worst not where there are few scanners, but where there are many reports and no clarity on what to grab first.

Who wins — AppSec teams with mature CI, where scans are already running and there's something to merge. Who loses — our assessment: not competitors, but the habit of opening a scanner report and reading top to bottom. The second risk is more serious: a "unified project map" looks convincingly neat right up until the first question about how complete it is. The article says nothing about that.

⚠️ What we don't know

The text contains not a single number about quality: how many false positives, on what dataset the model's resilience was measured, what attacks the AI Red Team generates, how much the product costs, and how it behaves on a monorepo with hundreds of services. The risk summary, as the vendor honestly writes, is "not a security assessment in a single number," and that's the right caveat. The problem is that the reader currently has no way to check where the line between the map and reality lies. Habr view counts are not a benchmark.

Prediction: within a quarter, at least one major AppSec vendor (Snyk, Checkmarx, or GitLab) will ship a built-in AI Red Team module or an LLM-interface attack mode. If that doesn't happen, the niche will remain with smaller players like INFERA.

Disclaimer: This material is a vendor article by INFERA Security about its own AI.SafeCode product, so the claimed attack surface mapping and AI Red Team capabilities have not been officially confirmed by independent testing, and the source contains no figures on false positives or check quality.

Source: Habr, INFERA Security blog