Before suspicious rows become model behavior.
The team wants a review queue for trigger-like strings, hidden instructions, unusual repetition, or source paths that could be useful to an attacker.
Data Poisoning
Surface adversarially useful rows and trigger patterns before they reach training data.
Use this when data could be intentionally shaped, externally supplied, or trigger-like enough that a security reviewer should inspect it before it reaches a fine-tune or evaluation set.
Security-relevant candidates, not verdicts.
Datascreen can surface rows and clusters that deserve review, preserve evidence, and record the reviewer decision without treating a candidate row as attack proof.
Source: Carlini et al., Poisoning Web-Scale Training Datasets is Practical, 2023; Souly et al., Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples, 2025; Zou et al., PoisonedRAG, 2024.
Watch walkthrough
Product demo: review trigger-like rows before fine-tuning.
Show suspicious rows grouped by trigger pattern and source context, then record whether the security reviewer removed, fixed, or escalated the rows.
What the product needs to make the decision obvious.
Trigger review
Group unusual strings, hidden controls, and repeated patterns that deserve security review.
Source trail
Keep where the row came from visible while the team decides what to do.
Human approval
Require a reviewer decision before acting on high-risk rows.
Security record
Export what was reviewed and what remains uncertain.
A prioritized list of rows and clusters that deserve human review.
The row, source, neighborhood, and reason shown together.
A record of what reviewers kept, removed, fixed, or escalated.
A workflow-ready handoff that states what was reviewed and what remains uncertain.