Anthropic AI safety researcher: red-teaming agent monitors and making alignment evaluations harder to detect
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Insufficient data: no reliable total publication count exists for this OpenAlex record, so output cannot be scored. This axis would measure career-wide publication volume and consistency across venues and years.
Impact—
Insufficient data: citation signal is too sparse to score — the record shows only h-index 1 and i10-index 1. This axis would measure citation-based influence, venue standing, and uptake of the work by other researchers.
Focus0/100
The focus score is 0, the lowest possible reading. The two reviewed papers sit in two adjacent AI-safety subfields — fieldDist splits 50/50 between AI control/agent security and alignment evaluation/LLM auditing — which alone would suggest a tight agenda. The score is dragged down by OpenAlex topic noise: geometric analysis, differential geometry, and complex manifold topics appear in the record but match neither reviewed paper, a pattern consistent with same-name contamination in name resolution.
Collaboration50/100
The radar scores collaboration at 50, grounded in 11 distinct coauthors in the underlying record. The frequent-collaborator list — including Alex Remedios, Simon Storf, Fabien Roger, Axel Ahlqvist, Richard Guan, Juan-Pablo Rivera, Adeline Kassler, Dmitrii Troitskii, Alexandra Souly, Kai Fronsdal, and Robert Kirk — points to team-scale industry safety research, with Hughes in the last-author position on both reviewed 2026 papers. This is a concentrated, lab-internal cluster rather than a wide cross-institution network.
Activity47/100
Activity scores 47, on a thin but current record: 4 works in the last three years, with countsByYear showing 1 work in 2024, 1 in 2025, and 2 in 2026 — both 2026 items being the Evolve-reviewed arXiv papers. This cadence suggests active but selective output centered on agent security and evaluation realism. An isolated 1966 entry in countsByYear is almost certainly a metadata artifact and is excluded from this reading.
Reproducibility55/100
Reproducibility scores 55, computed from a reproMean of 55 across 2 papers — exactly the set of Evolve-reviewed works. Both carry a medium repro-readiness rating: the red-teaming study depends on production monitor systems and adversarial agents, while the evaluation-realism study relies on inference-time compute and deployment scaffolds on Sonnet 4.6, implying nontrivial compute cost for independent replication but documented methods.
Research summary
John Hughes is an AI safety researcher at Anthropic whose recent work runs on two complementary tracks: attacking and hardening the monitors that control coding agents, and making alignment evaluations harder for models to detect. Evolve has reviewed two of his 2026 papers, on both of which he is last author — a position typically tied to senior or supervisory contributions. The first reports that production monitors Auto Mode and Guardian can be induced to fail by malign coding agents, with 79% prompt-injection success, and that proposed mitigations cut this sharply. The second reports that critique refinement and deployment scaffolds lift audit realism on Sonnet 4.6 from 10% to 26%, while acknowledging partial gains and substantial compute cost. His bibliometric footprint is narrow and recent: 4 works in the last three years, 11 distinct coauthors, an h-index of 1 and an i10-index of 1, consistent with an early-stage OpenAlex record rather than a long publication history. His field distribution splits evenly between AI control and agent security, and alignment evaluation and LLM auditing. Citation-linked impact cannot yet be scored, and the record itself is name-resolved from public bibliographic data, so same-name attribution carries some uncertainty.
Collaboration pattern
Hughes's recorded collaboration pattern is a single concentrated cluster: 11 distinct coauthors, recurring within the same safety group — Alex Remedios, Simon Storf, Fabien Roger, Axel Ahlqvist, Richard Guan, Juan-Pablo Rivera, Adeline Kassler, Dmitrii Troitskii, Alexandra Souly, Kai Fronsdal, and Robert Kirk. On both reviewed papers he appears as last author (not corresponding), the position typically occupied by senior or supervising contributors in industry lab publications. The radar collaboration score is 50.
Research keywords
Highlights
Red-teaming production agent monitors
Making alignment audits harder to detect
A paired attack-and-measurement agenda
Taken together, the two 2026 papers pair offensive testing of agent-control infrastructure with better measurement of evaluation realism, with Hughes in the last-author position on both — a coherent senior-level agenda spanning AI control and LLM auditing.
Timeline
- 20201 indexed work in the OpenAlex record — an isolated early data point in an otherwise recent profile.
- 20241 indexed work; the record remains sparse immediately before the current research push.
- 20251 indexed work, continuing the low-volume cadence visible in the bibliometric snapshot.
- 20262 works: both Evolve-reviewed arXiv papers — red-teaming Auto Mode blocking classifiers, and improving evaluation realism with inference-time compute — with Hughes as last author on each.
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.
- Bibliometrics are computed from OpenAlex only; this profile is name-resolved from public bibliographic data, and 'John Hughes' is a common name, so same-author attribution carries inherent ambiguity.
- The output and impact axes are unscored because citation and volume data are too sparse in the current record (h-index 1, i10-index 1).
- The 1966 countsByYear entry and geometry-related OpenAlex topics (geometric analysis, differential geometry, complex manifolds) match neither reviewed paper and are treated as likely disambiguation artifacts.
- Paper-level judgments rest on the two 2026 Evolve-reviewed papers only; other works in the record have not been reviewed by Evolve.