CMU AI-safety researcher working on post-hoc jailbreak defenses and insider-threat detection for LLM agents
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
6
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
Output metrics are not yet available in this snapshot: total publication and venue counts could not be resolved for this identity, so this axis would measure the volume, consistency, and venue mix of published works over time; it is left unscored rather than estimated.
Impact—
Impact data is insufficient in this snapshot: per-work citation counts and venue-level influence are not yet resolved for this OpenAlex identity, so no score is given. This axis would measure citations, citations per year, and uptake of the work by other researchers.
Focus0/100
The focus score is 0 because the cataloged work splits exactly 50/50 across two fields — LLM alignment robustness and LLM agent security — leaving no majority field. Both sit inside AI safety, so the split reflects two complementary safety sub-questions (defending refusal mechanisms; monitoring agent behaviour) rather than scattered interests. Topic counts (topic modeling 29, NLP techniques 12, adversarial robustness 11) reinforce an NLP-centred safety agenda.
Collaboration28/100
The collaboration score is 28, drawn from only 3 distinct coauthors in the snapshot. Recurring partners are Mona T. Diab and Virginia Smith, indicating small, stable, advisor-anchored teams rather than large consortia. Both cataloged 2026 papers are first-authored without a corresponding-author flag, consistent with an early-stage researcher working inside tight CMU-centered groups; the footprint is narrow but coherent.
Activity99/100
Activity scores 99, the strongest axis in this profile. Output ramps from 1 work in 2020 and roughly 2-5 per year through 2024 to 13 works in 2025 and 11 in 2026, with 29 works in the last three years. The near-doubling in 2025 and sustained 2026 pace indicate an intense, current publishing cadence centred on arXiv-first AI safety releases.
Reproducibility78/100
Reproducibility averages 78 across the 2 papers Evolve has assessed: MOLE (insider-threat detection) rates high readiness with its open 150-account benchmark, while Decoy Direction Optimization rates medium. Above-average readiness in both assessments suggests a habit of releasing usable artifacts, though the evidence base is limited to these two evaluations.
Research summary
Aashiq Muhamed is a researcher at Carnegie Mellon University working at the intersection of large language model safety and security, with a profile resolved from OpenAlex bibliographic data. The record shows a steep productivity ramp: from 1 work in 2020 to 13 in 2025 and 11 in 2026, with 29 works in the last three years — an activity score of 99, the strongest in this profile. Cataloged output splits evenly (50/50) between two AI-safety subfields: LLM alignment robustness and LLM agent security. Two first-authored 2026 arXiv papers anchor the catalog. Decoy Direction Optimization proposes a cheap post-hoc defense against abliteration-style refusal ablation by corrupting the attacker's contrastive direction estimate; Evolve judges it clever and cheap but notes it leaks under sustained adaptive re-estimation. MOLE contributes the first open insider-threat benchmark for AI agents — 150 accounts, 12 threat types — rated high for reproducibility readiness, though built on a single synthetic organization with sparse positives. Broader topic counts (topic modeling 29, NLP techniques 12, adversarial robustness 11) indicate NLP-rooted methods carried into safety applications. Collaboration is narrow: 3 distinct coauthors, with Mona T. Diab and Virginia Smith recurring. A reproducibility mean of 78 across 2 assessed papers suggests a genuine artifact-releasing habit; output and impact axes lack the citation data needed to score.
Collaboration pattern
The visible collaboration footprint is small: 3 distinct coauthors, with Mona T. Diab and Virginia Smith recurring most often. Both cataloged 2026 papers are first-authored (no corresponding-author flag), consistent with work inside small, advisor-anchored groups at Carnegie Mellon. The 28/100 collaboration score reflects this narrow network rather than isolation — the 2025-2026 output surge (24 works across the two years) has not yet widened the co-author pool in this snapshot.
Research keywords
Highlights
Cheap post-hoc defense against LLM abliteration
First open insider-threat benchmark for AI agents
Timeline
- 2020First work recorded in the OpenAlex snapshot (1 publication).
- 2021Output rises to 4 works, an early step-up from the 2020 baseline.
- 20245 works, closing a steady low-volume phase before the 2025 surge.
- 2025Publication pace jumps to 13 works, the highest annual count in the record.
- 202611 works to date, including both cataloged first-author arXiv papers: Decoy Direction Optimization and MOLE.
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.
- All bibliometrics come from an OpenAlex identity snapshot (identity: openalex); h-index 6 and i10-index 4 reflect only that record.
- Output and impact axes are unscored because per-paper citation and venue data were unavailable; both cataloged papers are arXiv preprints.
- The reproducibility mean of 78 rests on only 2 Evolve-assessed papers.
- The frequent-coauthor list contains a duplicate entry (Virginia Smith appears twice), so recurring partners are counted as two distinct people.