← Digests

Titus digest · filtered

AI-safety jobs — WFH / little travel / applied pivot

2026-09-15 · MS CS · applied AI eng · Anthropic bounty harness · no pubs · from Grok Bot

Voice picker uses whatever Safari exposes. Jamie Premium in iOS Spoken Content often does not appear here — that’s an Apple limit. If missing, pick another voice or use pre-recorded audio.

1 / 11
Ready.

TLDR (filtered)

Same survey as the full pack (~40 openings, seen 2026-09-15), re-filtered for: WFH / remote-US · little travel · applied software + AI eng · MS CS · AI-safety pivot · Anthropic model-safety bounty (harness built, no claim yet) · proposing testable ideas · no publications.

Keepers (5): #1 · Task Development Engineer · METR · #2 · Research Engineer / Senior Research Engineer · FAR.AI · #3 · Jailbreaking Lead, Red Team · FAR.AI · #4 · Evaluation & ML Systems Engineer · NVIDIA (AI Safety & Security) · #5 · Researcher / Senior Researcher · Epoch AI

Dropped: Anthropic hybrid evals/red-team (travel), FAR eng manager, SaferAI EU-preference, Cohere timezone tax, and all SF/London/UK-onsite honorable mentions.

Filter rules used

  • Logistics: WFH / remote-US first; little travel only (a few retreats/year OK; ≥25% office or SF/London mandatory = cut).
  • Background: years of software + applied AI; MS CS; daily agent/harness shipping is a plus.
  • Stage: pivoting into AI safety; bounty enrolled + harness; ideas that can be tested; no publications required.
  • Role shape: evals / red-team / control / ML systems / measurement eng > manager tracks > PhD-pubs research scientist seats.

#1 · Task Development Engineer · METR

remote worldwide evals / agents top fit

Why for Eric: Closest FTE/contract match to daily multi-agent harness + eval work. METR’s Time Horizons / agent capability evals are high-credibility independent safety measurement — and they hire remote with Pacific overlap (Chicago mornings work).

Requirements (public): Several years complex software eng; experience building hard (ideally agent-based) AI evals (RE-Bench, HCAST, SWE-bench Verified, Cybench, GPQA); Inspect preferred; Hawk / Time Horizons nice-to-have; high attention to detail.

Logistics: Remote worldwide; ~1–4h overlap with Pacific workday; 20–40h/wk flexible on contractor framing. America/Chicago is fine for overlap.

Travel: Not required for remote contractor framing; FTE path mentions Berkeley office culture / work trials — confirm.

Pay: Posted (careers / 80k Hours): $150–$300/hour contractor. Also seen on Lever (2026): FTE band $260,937–$385,490/year with Bay benefits/relocation language. Reconcile at apply time — page text recently shifted toward FTE-by-default while careers card still says remote contractor.

Day-to-day: Design novel hard tasks as model horizons grow; QA solvability; baseline/score; improve task-dev infrastructure.

Link: metr.org/careers · Lever posting

Caveats: Competitive; eval portfolio helps more than titles. Contract vs FTE ambiguity. Not a “lab insider” seat — independent evaluator impact.

#2 · Research Engineer / Senior Research Engineer · FAR.AI

remote global evals + red-team

Why for Eric: Explicitly ships pre/post-release adversarial evals of frontier models, novel attacks, robustness — builder/shipper safety without PhD-first framing. Remote + Berkeley optional.

Requirements (public, Research Engineer page historically): Implement ML algorithms, run experiments, analyze results; evals/red-teaming; open-source ML stack familiarity (PyTorch/HF). Senior bar = more ownership/scope.

Logistics: Remote and in-person Berkeley possible; hire remotely in most countries (80k Hours: Remote, global).

Travel: Work-related travel/equipment covered per older posting language; not “SF or nothing.”

Pay: Posted via 80,000 Hours (seen 2026-09-15): Senior Research Engineer $150–$250k. Research Engineer band not always listed on the card — treat as unknown / ask (older FAR page cited wide location-dependent ranges; do not invent).

Day-to-day: Frontier model adversarial evals; attack development; experiments on deception/robustness; code + paper co-authorship as MTS-style eng.

Links: far.ai/careers · RE Ashby 52e76732… · Senior RE 4f6fece8…

Caveats: Nonprofit/research org scale vs frontier-lab cash; Ashby pages are JS shells — read full JD in-browser. Competition still real among alignment-adjacent engineers.

#3 · Jailbreaking Lead, Red Team · FAR.AI

remote global red-team lead

Why for Eric: Direct line from enrolled Anthropic model-safety bounty + harness building into a paid red-team leadership seat at an org whose red team publishes jailbreak/finetune attack research affecting Claude/ChatGPT/Gemini safeguards.

Requirements: Mid (5–9y) experience per 80k Hours card; deep adversarial / jailbreak / eval execution (confirm full JD on Ashby).

Logistics: Remote, global.

Travel: Unknown beyond normal research org trips — assume low vs lab hybrid.

Pay: Posted (80k Hours): $170–$250k.

Day-to-day: Lead jailbreak/red-team technical strategy; evaluate frontier systems; turn findings into research + safeguard improvements.

Link: Ashby Jailbreaking Lead · board mirror 80k Hours · FAR AI

Caveats: Filter gate: “Lead” may want attack receipts — your bounty harness + one solid claim before apply strengthens this a lot. “Lead” title may expect prior published attacks or team lead proof — portfolio of authorized red-team results matters. Not the same as product pentest-only background.

#4 · Evaluation & ML Systems Engineer · NVIDIA (AI Safety & Security)

US remote (state list) evals / security tooling

Why for Eric: Evidence-first eval infrastructure for AI-powered vuln find/validate/patch — maps to shipping harnesses, metrics skepticism, agent measurement. Big-company stability + safety-adjacent cyber.

Requirements (public): Bachelor’s or equivalent + 5+ years ML eng/evaluation; designing benchmarks/metrics; solid Python for shared infra; experiment tracking/data pipelines. Preferred: security evaluation, agent/LLM behavior measurement, public benchmarks.

Logistics: US remote in listed locations (mirrors commonly: Santa Clara + remote CA/NC/NY/TN/FL). Texas not consistently listed on this exact JD — verify on jobs.nvidia.com before prioritizing.

Travel: Unknown; typical corp remote.

Pay: Posted base: $152,000–$241,500. ESTIMATE total comp: NVIDIA ML Eng Levels.fyi US medians often ~$200k–$330k+ TC depending on level (stock-heavy) — not this JD’s offer.

Day-to-day: Build benchmarking/reproducibility systems; define metrics/protocols; map every result to code+runs; keep conclusions reviewable.

Link: NVIDIA job 893396714449

Caveats: Filter gate: only keep if the live JD allows Texas remote (or you accept rare travel to a listed hub). State eligibility may block Tyler TX; applications may already be past “accept until July 30, 2026” language on some mirrors — confirm still open. Impact is AI-for-security tooling more than frontier alignment theory.

#5 · Researcher / Senior Researcher · Epoch AI

fully remote benchmarks / measurement

Why for Eric: Fully remote with PT–CET hiring; measurement/critique of AI progress fits “make numbers mean something.” Lower day-to-day red-team than #1–#3, but family-logistics excellent and credibility high.

Requirements: Open to varied backgrounds; researcher roles across multiple teams (incl. Benchmarking Reviews producing critiques of AI benchmarks). Prefer overlap with PT and UTC; can travel to ~3 retreats/year.

Logistics: Fully remote; many countries PT→CET. America/Chicago OK.

Travel: Prefer candidates who can attend ~3 staff retreats/year.

Pay: Posted examples: Researcher/Senior across teams often cited in wide bands (e.g. Benchmarking Reviews mirrors $100k–$200k); broader researcher postings sometimes higher — confirm per team. Do not assume top of range.

Day-to-day: Research trends/capabilities; write reviews/critiques; publish for policymakers and industry.

Links: Epoch Lever researcher · epoch.ai careers

Caveats: Filter gate: more writing/critique than harness shipping; no pubs is OK if you can show careful measurement write-ups. ~3 retreats/year is the travel ask. More research writing than agent harness shipping; retreat travel; not adversarial red-team primary.

Cut by this filter (still real jobs)

  • Anthropic · Research Engineer, Model Evaluations — Remote-friendly on paper but ≥25% SF/NYC office — fails little-travel filter.
  • Anthropic · Red Team Engineer, Safeguards — Hybrid office / travel required — closest thematic fit to bounty, but logistics fail.
  • FAR.AI · Engineering Manager, Red Team — People-manager bar; not an IC pivot seat.
  • SaferAI · Frontier AI Risk Management RE — Paris/London preference — not WFH-US first.
  • Cohere · MTS Safety for Agents — Remote-friendly but heavy UK/EU timezone overlap from Chicago.
  • Apollo / Redwood / OpenAI / xAI / UK AISI / NIST CAISI / METR Berkeley hybrid MTS — Already honorable-mentions: onsite or visa-heavy.

Full detail for cuts remains in the prior unfiltered digest if you want it later.

Suggested order (this filter)

  1. This week: one authorized Anthropic model-safety bounty submission via the harness (biggest free credibility upgrade; no pubs needed).
  2. Apply now: METR Task Development Engineer; FAR.AI Research Engineer (and Senior if the bar feels right).
  3. Stretch: FAR.AI Jailbreaking Lead — stronger after a bounty claim or two.
  4. If TX remote confirms: NVIDIA Evaluation & ML Systems (AI Safety & Security).
  5. Logistics-best research seat: Epoch AI (accept ~3 retreats/year; lean on careful write-ups, not a pub list).
  6. Parallel on-ramp: Anthropic Fellows and/or BlueDot while FTE pipes move.

On-ramps / not FTE (especially useful for this filter)

  • BlueDot Impact — Technical AI Safety (free course): bluedot.org
  • MATS research fellowships: matsprogram.org
  • ARENA technical bootcamps: arena.education
  • Anthropic Fellows (AI Safety / AI Security): 4 months FT; remote-friendly US/UK/Canada with work auth; stipend $3,850/week USD (+ ~$15k/mo compute). On-ramp with historically strong conversion to Anthropic safety FTE. Fellows Greenhouse
  • Anthropic model safety bug bounty + red.anthropic.com — already enrolled; finish a clean claim.
  • xAI/SpaceX HackerOne (authorized only): hackerone.com/x
  • 80,000 Hours map + board: 80000hours.org/ai · jobs.80000hours.org
  • METR General Expression of Interest (flexible): metr.org/careers
  • Hawk open eval stack: hawk.metr.org — practice Inspect-like task work publicly.

For this profile, prioritize next: (1) file one Anthropic model-safety bounty claim with the harness you already built — that is the cleanest credibility upgrade with no pubs; (2) BlueDot Technical AI Safety if you want shared vocabulary fast; (3) Anthropic Fellows (remote-friendly US, paid, 4 months) as a structured pivot if FTE timing is slow; (4) METR EOI even while applying to Task Dev Eng.

Sources & honesty

Survey date 2026-09-15. Pay marked Posted vs ESTIMATE. Confirm live JDs before investing hours. This page is a filter of the earlier pack, not a new scrape.