# Research

Independent technical reports by Anish Prashun on AI agent reliability, security and privacy. Each is archived on Zenodo with a DOI and comes with open code, synthetic data and one command that reproduces every number in it. Papers and data: CC BY 4.0. Code: MIT.

## Can Several Minimal Findings Reveal a Hidden Attribute?

An Exactly Solvable Pilot on Linked, Coarsened and Task-Output Releases, with a Release-Audit Toolkit

- Author: Anish Prashun (independent researcher)
- Published: 2026-11-17
- Page: https://www.anishprashun.me/research/minimal-findings-hidden-attribute
- Paper (PDF): https://www.anishprashun.me/research/minimal-findings-hidden-attribute.pdf
- Code and data: https://github.com/Anishpras/linked-findings-inference
- DOI: https://doi.org/10.5281/zenodo.24304575
- Reproduce: `python3 run_pilot.py`

Multi-stage AI workflows often pass a few structured findings between stages instead of raw records. This study measures how much several such findings, combined, reveal about an attribute that none of them states, in a synthetic testbed where the best possible attacker is computed exactly. It ships `release_audit`, a standard-library toolkit that computes the same leakage budget and the best release designs for your own pipeline.

Key findings:

- Linking two releases about the same household raised attacker accuracy to 60.0%, against 51.9% and 53.3% for either release alone (38.9% for always guessing the most common band).
- In 193 of 3,000 test households the linked releases ruled out the low reserve band with certainty, which neither release allows on its own.
- Coarsening the linked releases removed 1.8 points of accuracy while keeping 80% of report action items, and met its pre-set criterion in only 3 of 6 seed runs.
- Releasing only the report-checklist answers kept full report utility but still leaked more than either single release.

Cite: Prashun, A. (2026). Can Several Minimal Findings Reveal a Hidden Attribute? An Exactly Solvable Pilot on Linked, Coarsened and Task-Output Releases, with a Release-Audit Toolkit [Technical report]. Zenodo. https://doi.org/10.5281/zenodo.24304575

```bibtex
@misc{prashun2026findings,
  author    = {Prashun, Anish},
  title     = {Can Several Minimal Findings Reveal a Hidden Attribute? An Exactly Solvable Pilot on Linked, Coarsened and Task-Output Releases, with a Release-Audit Toolkit},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24304575},
  url       = {https://doi.org/10.5281/zenodo.24304575},
  note      = {Technical report, code and data}
}
```

## Safe Retry After an Ambiguous Tool Timeout

A Hand-Checkable Conformance Suite for Agent Recovery with Provider Contracts and Webhook Evidence

- Author: Anish Prashun (independent researcher)
- Published: 2026-11-17
- Page: https://www.anishprashun.me/research/safe-retry-ambiguous-tool-timeout
- Paper (PDF): https://www.anishprashun.me/research/safe-retry-ambiguous-tool-timeout.pdf
- Code and data: https://github.com/Anishpras/ambiguous-retry-recovery
- DOI: https://doi.org/10.5281/zenodo.24304849
- Reproduce: `python3 run_sim.py`

When an AI agent's side-effecting tool call times out, the agent cannot tell whether the action happened. This deterministic, hand-checkable simulator scores retry policies against stale status reads and webhooks that arrive late, twice, out of order or not at all, under providers with and without idempotency-key deduplication. It ships a provider contract checklist, a decision table and commands to test your own scenarios and policies.

Key findings:

- Across 17 scripted operations without provider deduplication, immediate retry duplicated 12 and a single status check duplicated 8; the evidence-reconciling policy duplicated 1, only where the provider broke its own guarantee.
- Stable idempotency keys removed every policy-caused duplicate, and rotating the key on each retry brought them all back.
- Webhook delay, duplication and reordering did not affect a reconciler that only adds facts per attempt, while ignoring webhooks left 41 of 120 timing cases unresolved instead of 16.
- Wait-then-verify is safe exactly when its wait is at least the longest pending time plus the status lag, minus the timeout.

Cite: Prashun, A. (2026). Safe Retry After an Ambiguous Tool Timeout: A Hand-Checkable Conformance Suite for Agent Recovery with Provider Contracts and Webhook Evidence [Technical report]. Zenodo. https://doi.org/10.5281/zenodo.24304849

```bibtex
@misc{prashun2026retry,
  author    = {Prashun, Anish},
  title     = {Safe Retry After an Ambiguous Tool Timeout: A Hand-Checkable Conformance Suite for Agent Recovery with Provider Contracts and Webhook Evidence},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24304849},
  url       = {https://doi.org/10.5281/zenodo.24304849},
  note      = {Technical report, code and data}
}
```

## What Evidence Do AI Agent Audit Logs Need After Redaction?

A Matched-Pair Pilot on Verifying Consent from Masked and Aliased Logs

- Author: Anish Prashun (independent researcher)
- Published: 2026-11-12
- Page: https://www.anishprashun.me/research/audit-logs-after-redaction
- Paper (PDF): https://www.anishprashun.me/research/audit-logs-after-redaction.pdf
- Code and data: https://github.com/Anishpras/redaction-audit-pilot
- DOI: https://doi.org/10.5281/zenodo.24163831
- Reproduce: `python3 run_pilot.py`

AI agent audit logs often have to be redacted before a reviewer sees them, yet the reviewer must still decide whether each recorded action was authorised. This matched-pair pilot measures which evidence a redacted log must keep for that decision. It ships `redaction_audit`, a standard-library toolkit that tests your own redaction schemes against your own authorisation policy.

Key findings:

- Uniform masking made 6 of 10 matched pairs byte-identical, and the reviewer could confirm none of the 10 authorised actions.
- Consistent within-run aliases restored all 20 verdicts while exposing none of the 66 planted raw values.
- Removing expiry or revocation evidence left 12 of 20 decisions unresolved, including every authorised action.
- Hashing identifiers with a fresh salt per event led a reviewer unaware of the scheme to flag all 10 authorised actions as violations, so redacted logs should declare how they were redacted.

Cite: Prashun, A. (2026). What Evidence Do AI Agent Audit Logs Need After Redaction? A Matched-Pair Pilot on Verifying Consent from Masked and Aliased Logs [Technical report]. Zenodo. https://doi.org/10.5281/zenodo.24163831

```bibtex
@misc{prashun2026redaction,
  author    = {Prashun, Anish},
  title     = {What Evidence Do AI Agent Audit Logs Need After Redaction? A Matched-Pair Pilot on Verifying Consent from Masked and Aliased Logs},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.24163831},
  url       = {https://doi.org/10.5281/zenodo.24163831},
  note      = {Technical report, code and data}
}
```
