AI security researcher at Microsoft: agent incident reporting and inference-time LLM backdoor detection
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
This axis would measure publication volume and productivity over time, but the underlying counts are not yet available in this snapshot. With only two Evolve-reviewed papers on record, no output score is computed.
Impact—
This axis would quantify citation-based influence such as citation counts and h-index, but those metrics are not available in the current data. No impact score is estimated here.
Focus0/100
The cataloged record splits evenly — 0.5 to AI security and governance and 0.5 to LLM security (backdoor detection / defense) — which yields a focus score of 0 on this snapshot. Read charitably, the two halves are adjacent security subfields around agent and model risks, so the spread reflects a deliberately broad security agenda rather than scatter across unrelated areas, based on two reviewed papers.
Collaboration68/100
A collaboration score of 68 rests on 29 distinct coauthors in the snapshot, including recurring partners such as Battista Biggio, Somesh Jha, Eugene Bagdasarian and Luca Beurer-Kellner. Both 2026 catalog papers are multi-author efforts with the researcher in a middle authorship position rather than leading solo, consistent with collaborative industrial research across teams and institutions.
Activity—
This axis would track recent publication cadence and year-over-year research activity, but the underlying temporal data is insufficient in this snapshot. No activity score is computed.
Reproducibility38/100
Reproducibility averages 38 across the two Evolve-reviewed papers. SpecGuard is rated medium readiness but its review notes no released code, while the agent incident-reporting framework is rated low because it is a position paper without solicitation methods or validation. The pattern suggests strong conceptual framing not yet matched by artifact release.
Research summary
Andrew Paverd is a researcher at Microsoft (United States) working at the intersection of AI security and governance and LLM security. This Evolve profile is a full-tier, name-resolved record built from public bibliographic data linked to ORCID 0000-0003-2188-5285, combined with two Evolve-reviewed papers, both posted to arXiv in 2026. The first, 'SpecGuard: Inference-Time Backdoor Detection For Free', proposes detecting backdoored LLMs at inference time using accept/reject traces from speculative decoding — essentially repurposing an existing decoding mechanism as a free runtime monitor. Evolve's review credits it with broad, strong detection results while noting that naturalistic payloads can evade it and that no code has been released. The second, 'Beyond Predictable Paths', is an expert position paper arguing that security incident reporting must be redefined for AI agents, incorporating trajectories, autonomy, delegation and multi-agent data; the framework is clear but lacks a solicitation method or validation. The snapshot splits evenly (0.5/0.5) across the two fields, records 29 distinct coauthors with a collaboration score of 68, and yields a reproducibility score of 38 averaged over two papers. Output, impact and activity axes are unscored for lack of data. The overall picture is an industrial researcher mapping the lifecycle of agent and model security — from runtime detection mechanisms to incident-reporting institutions — in collaborative, multi-author teams.
Collaboration pattern
The recorded work is highly collaborative: the snapshot lists 29 distinct coauthors and a collaboration score of 68. Recurring partners include security researchers such as Battista Biggio, Somesh Jha and Eugene Bagdasarian, alongside other frequent coauthors including Luca Beurer-Kellner, Daphne Ippolito and Heather Frase. On both cataloged 2026 papers he sits in a middle authorship position and is not the corresponding author, indicating contribution within larger teams rather than solo-led projects.
Research keywords
Highlights
Free inference-time backdoor detection via speculative decoding
A blueprint for reporting AI agent security incidents
Detection and disclosure: two ends of one LLM security lifecycle
Taken together, the two 2026 papers sketch a lifecycle view — runtime detection of compromised models on one end, structured incident reporting for autonomous agents on the other — positioning the author across both technical mechanism and governance process.
Timeline
- 2026Co-authored 'SpecGuard: Inference-Time Backdoor Detection For Free' (arXiv), proposing backdoor detection from speculative-decoding traces; Evolve rates its reproducibility readiness medium.
- 2026Co-authored 'Beyond Predictable Paths: Redefining AI Security Incident Reporting for Agents' (arXiv), an expert position paper on agent incident reporting; Evolve rates its reproducibility readiness low.
- 2026Evolve full-tier snapshot records 29 distinct coauthors, an even 0.5/0.5 split across two security fields, and collaboration/reproducibility scores of 68 and 38.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric dimensions are computed from OpenAlex-based public bibliographic data; the output, impact and activity axes lack sufficient underlying data and are intentionally unscored.
- Qualitative assessments rest on only two Evolve-reviewed arXiv papers (both 2026); peer-reviewed venue acceptance is not captured in this snapshot.
- Reproducibility (mean 38 over two papers) is an early, small-sample signal rather than a stable estimate.
- Identity is resolved via the provided ORCID (0000-0003-2188-5285); no biography, award or career information beyond the input data is included.