id stringlengths 33 33 | messages listlengths 9 97 | metadata dict |
|---|---|---|
agent_task_20260610T221810777696Z | [
{
"role": "system",
"content": "You are a software engineering assistant with full bash terminal access inside a Docker container.\n\n## Action format\nEvery response must include exactly one bash tool call.\nReasoning text is allowed, but the bash tool call is what executes actions.\n\n## Rules\n- Director... | {
"project": "assimp",
"task_mode": "patch",
"vulnerability_id": "vulnerability_0",
"validation_status": "success",
"teacher_regime": "faithful",
"include_hints": false,
"api_calls": 12,
"total_steps": 12,
"generation_method": "fuzz_guided",
"model_id": "gpt-5.4-mini"
} |
agent_task_20260610T220852479555Z | [
{
"role": "system",
"content": "You are a software engineering assistant with full bash terminal access inside a Docker container.\n\n## Action format\nEvery response must include exactly one bash tool call.\nReasoning text is allowed, but the bash tool call is what executes actions.\n\n## Rules\n- Director... | {
"project": "assimp",
"task_mode": "poc-san",
"vulnerability_id": "vulnerability_1",
"validation_status": "success",
"teacher_regime": "faithful",
"include_hints": false,
"api_calls": 8,
"total_steps": 8,
"generation_method": "fuzz_guided",
"model_id": "gpt-5.4-mini"
} |
agent_task_20260610T223850533742Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"assimp","task_mode":"patch","vulnerability_id":"vulnerability_1","validation_status":"su(...TRUNCATED) |
agent_task_20260610T231107696644Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"assimp","task_mode":"patch","vulnerability_id":"vulnerability_2","validation_status":"su(...TRUNCATED) |
agent_task_20260611T000007816808Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"jq","task_mode":"patch","vulnerability_id":"vulnerability_0","validation_status":"succes(...TRUNCATED) |
agent_task_20260611T002013007113Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"jq","task_mode":"patch","vulnerability_id":"vulnerability_2","validation_status":"succes(...TRUNCATED) |
agent_task_20260610T232435283034Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"brotli","task_mode":"poc-san","vulnerability_id":"vulnerability_1","validation_status":"(...TRUNCATED) |
agent_task_20260610T232815409342Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"brotli","task_mode":"patch","vulnerability_id":"vulnerability_1","validation_status":"su(...TRUNCATED) |
agent_task_20260611T011923549702Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"brotli","task_mode":"patch","vulnerability_id":"vulnerability_2","validation_status":"su(...TRUNCATED) |
agent_task_20260611T021155706185Z | [{"role":"system","content":"You are a software engineering assistant with full bash terminal access(...TRUNCATED) | {"project":"assimp","task_mode":"patch","vulnerability_id":"vulnerability_3","validation_status":"su(...TRUNCATED) |
Vuljector SEC-bench Teacher Trajectories
1083 faithful (no-hint) success trajectories of an agent solving SEC-bench
vulnerability tasks. Every trajectory is from the gpt-5.4-mini teacher in the
faithful regime (include_hints=false) — no other model and no hint-assisted runs
are included. Natural-language reasoning is captured. Each row is one trajectory:
messages (system/user/assistant/tool, with reasoning + bash/cmd tool calls on
assistant turns) and metadata. Secret-scrubbed. All trajectories are in a single file
(all_trajectories.jsonl, train split); the source pipeline of each row is recorded
in metadata.generation_method.
Vulnerabilities come from three generation pipelines:
| generation_method | patch | poc-san | total |
|---|---|---|---|
fuzz_guided |
537 | 217 | 754 |
workflow_rag |
136 | 12 | 148 |
workflow_llm |
76 | 15 | 91 |
cve_reversal |
90 | 0 | 90 |
| TOTAL | 839 | 244 | 1083 |
- fuzz_guided — bug injected at a location reachable by an existing fuzzer harness.
- workflow_ — bug from a RAG or LLM candidate-generation workflow.
- cve_reversal — task reconstructed from a real OSS-Fuzz CVE testcase replay (patch-only; no synthetic sanitizer report, so no poc-san variant).
71 projects. Two task modes: patch (produce a fix) and poc-san (produce a
crashing PoC verified by the sanitizer).
poc-san is quality-filtered. SEC-bench's poc-san repro can crash on a baked-in
reference seed even when the submitted artifact has no real PoC, which inflates
"success". Every poc-san trajectory here was verified to have a real selected_fuzzer
and a genuine non-empty submitted PoC file (empty / description-only / diff-only
submissions were dropped). patch trajectories are scored on the real git_patch, so
they are unaffected.
Row schema
{
"id": "agent_task_...",
"messages": [ {"role": "system|user|assistant|tool", "content": "...",
"reasoning": "...", "tool_calls": [...]} ],
"metadata": {
"project": "...", "vulnerability_id": "...", "task_mode": "patch|poc-san",
"model_id": "gpt-5.4-mini",
"validation_status": "success", "teacher_regime": "faithful",
"include_hints": false, "api_calls": 0, "total_steps": 0,
"generation_method": "fuzz_guided|workflow_rag|workflow_llm|cve_reversal"
}
}
- Downloads last month
- 84