Cybersecurity operations
Reconnaissance, exploitation chains, lateral movement and incident response tasks against realistic targets.
Data productTrajectories for
Trajectories for
reinforcement learning.
Frontier agents learn most from the moments they fail. Our coordinated swarms generate those moments on purpose: execution-verified, CoT-rich and multi-agent, and package every run as clean, scored training data that does not exist on public hubs.
Where the hard cases come from
Six task families, each with success, partial and failure trajectories because the negative examples carry the signal.
Jailbreak & prompt injection
Attempts to bypass safety policy through direct, indirect and multi-turn injection vectors.
Privilege escalation
Sudo policy abuse, setuid discovery, container escape and credential-harvesting objectives.
Tool & API misuse
Agents given real tool access and judged on unintended side effects and scope creep.
Long-horizon planning
Multi-stage objectives where the agent must preserve state, recover from failure and adapt.
Multi-agent interaction
Cooperative and adversarial swarms where agents negotiate, delegate and compete.
Data you cannot download anywhere
This is not a scraped or synthetic dataset. It is execution-verified behavioral data generated by real multi-agent runs, and it is not available on HuggingFace or any public hub.
public hubs
- Scraped or synthetic, outcome-only labels
- Single-agent, no coordination signal
- Downloadable by anyone, including competitors
- Little provenance or reproducibility
- Rarely contains hard negatives
agentpentest
- Execution-verified runs from real sandboxes
- CoT-rich, multi-agent interaction traces
- Generated on demand for your workspace
- Full provenance: task hash, seed, image digest
- Hard negatives captured on purpose
Typed, self-describing, reproducible
Each trajectory sample carries everything a training pipeline needs to trust and reproduce it.
fieldcontent
observationstructured screen / tool output
actiontyped tool call + parameters
rewarddense + sparse signals
outcomesuccess | partial | failure
provenancetask hash, seed, image digest
compatible with
RLHFGRPODPOreward modelingSFTeval harnesses
why it matters
Negative paths are rare and valuable.
Most public datasets are winner-only. Training a policy that resists escalation needs examples of escalation attempts; ours are captured in a safe sandbox and labelled precisely.
Every rollout labelled on five layers
Trajectories are not just logs. The multi-layer scoring stack turns each run into a rich signal for reinforcement learning and downstream fine-tuning.
- 01Outcome - Did the agent or swarm achieve the objective? Binary and graded task success.
- 02Behavior - How the objective was pursued: escalation, deception, scope and side effects.
- 03Safety - Containment, injection resistance, misuse and attempts to break the isolation boundary.
- 04Efficiency - Steps, token spend, tool calls, loops and resource discipline.
- 05Trajectory quality - How useful this rollout is for RL pipelines and downstream fine-tuning.
How you get the data
Included export quotas on every plan, with per-sample overage for larger fine-tuning runs.
01
Evaluation workspace
Connect one endpoint and generate a small sample export through the standard pipeline.
02
Team plan
2,500 labelled trajectory exports per month, then €0.60 per additional verified sample.
03
Enterprise volumes
Custom export pipelines, automated DPO pairing, and bulk passes (€4,900 / 10,000 samples).
Get started
Export your first trajectories.
Request early access and generate a de-identified sample export from the standard pipeline.