macOS Threat Investigation
Mar 2026BlueBench-Intrusion-001: Real macOS infostealer intrusion spanning incident response, threat hunting, and detection engineering
| Rank | Models (9) | Acc | Cost | Lat |
|---|---|---|---|---|
| 1 |  | 86.8% | $2.81 | 36m 49s |
| 2 |  | 84.2% | $3.90 | 48m 38s |
| 3 | | 81.6% | $3.33 | 37m 5s |
| 4 | | 75.0% | $4.94 | 43m 34s |
| 5 | GLM-5 | 71.0% | $3.76 | 62m 7s |
| 6 |  | 68.8% | $0.86 | 41m 3s |
| 7 |  | 68.4% | $0.36 | 29m 38s |
| 8 | | 59.2% | $0.42 | 5m 15s |
| 9 | | 54.0% | $0.29 | 88m 53s |
About This Benchmark
Sample Questions
Q: What persistence artifact was installed? Provide the full plist path.
A: /Library/LaunchDaemons/com.*****.plist (redacted)
Q: What built-in utility is used to decode content immediately before the main payload executes?
A: base64 -d
Q: Write and validate a correlation query that ties local credential validation to a subsequent privileged action within a short time window.
A: [SQL query — executed against the live dataset to validate detection coverage]
Methodology
Scoring
- Accuracy: LLM-judged 0–1 per question using a weighted rubric: technical accuracy (60%), completeness (25%), specificity (15%)
- Cost: USD per task based on token usage
- Latency: Wall-clock time to complete each investigation track
- Completion: Percentage of tracks finished without unrecoverable errors
Setup
- Logs from a real macOS intrusion loaded into a sandboxed query environment
- Agents given SQL-based query tools to investigate 416K+ events across 14 log sources
- 36 scored questions across three investigation tracks: incident response, threat hunting, and detection engineering
- Questions range from alert triage and kill-chain reconstruction to containment decisions and detection rule authoring
Controls
- Same minimal system prompt for all models, no per-model tuning
- "Thinking" mode enabled where available
- Agent loops capped at 40 iterations (tools removed to force an answer)
- All models given identical tool access and data
Key Findings
Accuracy (All Tracks)
GPT-5.4 led at 87%, followed by Opus 4.6 (84%) and GPT-5.3 Codex (82%). GPT-5.4 was the most consistent across tracks (83-92%), while Gemini 3.1 Pro scored 89% on threat hunting but dropped to 57-58% elsewhere.
Accuracy by Model
Track Breakdown
Each track rewarded different strengths. IR was hardest (top score 83%). Gemini 3.1 Pro dominated threat hunting at 89%. GPT-5.4 and Opus 4.6 tied at 92% on detection engineering. One question stumped every model: locating the hidden credential cache path.
Cost (All Tracks)
IR was the most expensive track across models. Kimi K2.5 had the best value at $0.36/task for 68% accuracy, roughly 8x cheaper than GPT-5.4. GPT-5.4 Mini hit $0.23/task on detection engineering while still scoring 71%.
Cost per Task
Speed (All Tracks)
GPT-5.4 Mini finished tracks in ~5 minutes. GPT-5.4 and GPT-5.3 Codex averaged ~37 minutes. Gemini 3.0 Flash was slowest at 89 minutes per track.
Task Duration (avg)
Reliability (All Tracks)
Eight of nine models achieved 100% completion. Gemini 3.1 Pro had 2 failures out of 40 runs (95%).
Task Completion Rate
Model Recommendations
- GPT-5.4 — Best for high-stakes investigations. Highest accuracy at 87% with no weak tracks (83-92%) at ~$2.81/task.
- Kimi K2.5 — Best value overall. Achieves 68% accuracy at ~$0.36/task with 100% reliability, and scores 79% on threat hunting specifically.
- GPT-5.4 Mini — Best for rapid triage. Completes tracks in ~5 minutes and scores 71% on detection engineering at just $0.23/task.
- Gemini 3.1 Pro — Best for threat hunting. Scores 89% on that track (highest of any model on any track) at $1.63/task, but weaker on IR and detection engineering.
- Claude Opus 4.6 — Second highest accuracy at 84%, tied for best on detection engineering (92%). Strongest pick in the Anthropic ecosystem.
Caveats
- Analysis questions use LLM-judged scoring, which introduces some variability compared to exact-match evaluation.
- Costs are measured per task (12 tasks per track).
- Gemini 3.1 Pro was in preview at the time of evaluation with tight rate limits and high API response latency. Its wall-clock times may not reflect production performance.
Other Benchmarks
Real CTF challenges from CSAW competitions covering reverse engineering, forensics, and miscellaneous problem-solving
Defensive security CTF challenges testing forensics, reverse engineering, and miscellaneous security skills
Blue team CTF scenarios testing incident response and threat hunting
Multi-label classification of MITRE ATT&CK tactics and techniques from Sigma rules
Multiple-choice cybersecurity knowledge evaluation across 10,000 questions
AI for the blue team.
Run Cotool's harness in your environment to get real security work done