pretty_name: UOF Agent Boundary Policies v2
language:

• en
  license: cc-by-4.0
  tags:
• unicode
• synthetic-data
• security-research
• evaluation

UOF Agent Boundary Policies v2

Synthetic boundary-policy cases across four parser surfaces: plain_text, markdown, json_string, and html.

The six techniques are word_joiner, zero_width_join, bidi_isolate, unicode_escape, html_entity, and quoted_boundary. Each row keeps visible text, an escaped hidden payload, the decoded projection, decode steps, a policy decision, a risk band, and a defensive handling instruction. All rows are marked as synthetic harmless payloads.

Release contract

• Generator: uof-frontier-v2.0.0
• Seed: 20260915
• Rows: 720 (train 420; validation 60; test 120; ood 120)
• Split isolation: the publication gate found no duplicate IDs and no root overlap across splits.
• Policy decisions are balanced across allow, preserve_literal, review, and reject.

Limitations

This is a synthetic parser-boundary policy set, not a complete prompt-injection corpus and not an agent firewall. The payloads are harmless examples for inspecting boundaries; they are not real attacks or evidence that a specific agent framework will decode the same way. Keep decoding, display, authorization, and human-confirmation controls outside the model.

Load

from datasets import load_dataset
ds = load_dataset("NewSonnet/uof-agent-boundary")

License

CC BY 4.0.