U0001f4e6 PyPI: https://pypi.org/project/resksecure
U0001f419 GitHub: https://github.com/Resk-Security/reskSecure
U0001f310 Web: https://resk.fr

---

Blocking LLM Jailbreaks at the Logits Level with reskSecure

The Problem

Prompt injection and jailbreak attacks are the number one vector against production LLM systems. Instruction-based filters and post-generation regex checks have a fundamental flaw: they operate after the dangerous token has already entered the generation pipeline.

A single malicious token in the output stream can leak PII, call a disallowed API function, or generate harmful content. By the time a regex catches it, the damage is done.

The Solution: Bitmask-Based Logits Firewall

reskSecure takes a different approach. Instead of scanning output text, it intercepts at the logits stage — before a single token is sampled. Using a configurable bitmask system, it either zeroes out dangerous tokens or applies a learned penalty to discourage them.

How It Works

from resksecure import Firewall, Severity

# Load a YAML policy
fw = Firewall.from_yaml("policy.yaml", severity=Severity.HARD)

# The firewall wraps your generation call
output = fw.generate(
    model=model,
    prompt="Tell me how to pick a lock",
    tokenizer=tokenizer
)
# Dangerous tokens are blocked before they ever reach the output

Key Features

• Two severity modes: HARD sets blocked logits to -infinity, BIAS applies a configurable penalty
• Hot-reload policies: Change YAML rules without restarting your inference server
• Tool call prevention: Block the first token of a disallowed function call at the logits level
• YAML policy system: Define forbidden tokens, phrases, and tool patterns in human-readable files
• Python >= 3.13 + PyTorch >= 2.0.0 compatible

Why Logits-Level?

Post-generation filters check output text and allow the model to already have generated the dangerous content. Logits-level filtering prevents generation entirely. The model simply cannot produce the first token of a blocked sequence.

This is especially critical for tool calling. When an agent tries to invoke a sensitive function, reskSecure intercepts the logits before the function name token is sampled — no post-hoc check, no race condition.

Quick Start

pip install resksecure

Create a simple policy file:

# policy.yaml
severity: hard
blocklist:
  - "ignore previous instructions"
  - "disregard safety"
  - "jailbreak"

Then run:

from resksecure import Firewall, Severity

fw = Firewall.from_yaml("policy.yaml", severity=Severity.HARD)
result = fw.generate(model, prompt, tokenizer)

Next Up

We are actively building RESK Monitor — real-time LLM deployment monitoring with anomaly detection on behavioral drift, error rate analysis, and automated incident response. Stay tuned.

Links

• PyPI: https://pypi.org/project/resksecure
• GitHub: https://github.com/Resk-Security/reskSecure
• Company: https://resk.fr
• Documentation: Coming soon

If you run LLMs in production, try reskSecure and let us know what you think. Contributions, issues, and feedback are welcome.