gemma-coder — security-hardened coding model

gemma-coder is a QLoRA fine-tune of google/gemma-4-26B-A4B-it (a mixture-of-experts base with ~4B active parameters) that serves as the default agent model for the remote-agent-dev-platform. It runs inside a sandboxed agent that builds, executes, and tests everything it writes.

The base model already writes competent code across the six languages this agent uses. Correctness is no longer the frontier — safety is. So the model's objective has been re-scoped from "write working code" to write code that does not introduce vulnerabilities, and recognise the ones already in a codebase. The measuring stick for that objective is ExploitGym.

Why ExploitGym

ExploitGym (Berkeley RDI, with the Max Planck Institute for Security and Privacy, UC Santa Barbara, Arizona State, Anthropic, OpenAI, and Google; arXiv 2605.11086) is a benchmark of 898 real-world vulnerabilities across three domains:

Domain	Instances	Source
Userspace C/C++ (FFmpeg, OpenSSL, …)	520	OSS-Fuzz / OSV
V8 JavaScript engine (Chromium)	185	real V8 bugs
Linux kernel	193	privilege-escalation tasks

Each task ships vulnerable source, a proof-of-vulnerability input, and a containerised runtime with toggleable mitigations (ASLR, stack canaries, the V8 sandbox, KASLR). ExploitGym measures whether an agent can turn a known bug into a working exploit — the hardest possible probe of whether a model truly understands a vulnerability rather than pattern-matching its surface.

We use it defensively. Every ExploitGym instance is a labelled, executable example of a real defect and exactly what it takes to trigger it. That is the richest possible curriculum for the two things this model is being aligned to do:

1. Not emit those defect classes in the first place — the exact memory-safety, type-confusion, and privilege-boundary mistakes ExploitGym is built from become negative examples the model learns to avoid as it writes.
2. Find them in existing code — given a diff or a file, locate the vulnerability, name its class (CWE), explain the trigger condition, and propose a fix, evaluated against ExploitGym's ground-truth bugs and its agent-as-a-judge verifier.

The benchmark is used strictly as an evaluation and alignment target for defensive security (secure generation, vulnerability detection, and code review) — not to produce exploits. ExploitGym's own runs used structured, gated security-research access; this model does not reproduce that.

Intended uses

• Secure code generation across Python, JavaScript/React, Go, Java, and Swift, inside a sandbox that executes and tests every output.
• Vulnerability detection and code review — flagging insecure patterns, naming the CWE class, and proposing fixes as it or a human writes.

Not intended for: producing exploits or offensive tooling; safety-critical systems; or running generated code unreviewed. This is a small, free-tier-trained model — treat every output as a draft a human must review and test.

Status & roadmap

• Shipped today: the six-language coding QLoRA described under Training below, with the coding evaluation shown.
• In progress: re-scoping toward the secure-coding + vulnerability-detection objective above, evaluated against ExploitGym. Security-hardened results are not yet established — this card states the objective, not a completed benchmark. It will be updated with ExploitGym-measured numbers once they exist; until then, no ExploitGym score should be attributed to this model.

Training

• Method: QLoRA (Unsloth), 4-bit, LoRA rank 8 / alpha 16 on attention layers, lr 2e-5, max seq len 768, AdamW-8bit.
• Compute: weekly 8-hour sessions on Kaggle's free dual-T4 GPUs, resumed across cycles.
• Coding data: bigcode/self-oss-instruct-sc2-exec-filter-50k.

Evaluation

Current coding evaluation — multi-language sandboxed pass@1 over executable test suites (promotion threshold 80%):

Language	pass@1
Go	4/4
Java	4/4
JavaScript	4/4
Python	4/4
Swift	skipped (no toolchain in the eval image)

This is a small smoke-test suite (16 executed problems), not a broad coding benchmark — read it as a promotion gate, not a capability claim. The security objective above is evaluated separately against ExploitGym and is not yet reported.

Limitations

• Small, free-tier-trained MoE fine-tune; it can produce incorrect or insecure code. Always review and test before use.
• The security re-scoping is in progress; do not rely on this model for vulnerability detection until ExploitGym-measured results are published here.
• ExploitGym is referenced as a defensive evaluation target only. Do not use this model to generate exploits.