CV | Bobby Filar

Experience

2022 — present

Sublime Security

Designed and co-developed ADÉ (Autonomous Detection Engineer), a specialized coding agent that writes detections through a knowledge base, tool use, and specialized sub-agents.
Co-authored an agent evaluation framework for LLM-generated security rules with hold-out human baselines; introduced metrics for detection accuracy (unique-TP precision), brittleness → robustness, and cost-to-pass syntactic validation.
Built MQLBench, a 30k-example natural-language-to-security-DSL benchmark with a four-metric evaluation framework (validity, field-overlap, LLM-judge, execution-grounded behavioral scoring) and a public six-model leaderboard.
Architected ASA (Autonomous Security Analyst), a deep-reasoning agent for triaging phishing emails.
Designed and shipped a multi-stage prompt-injection detection system (ModernBERT classifier + conformal calibration + LLM-judge cascade).
Designed a safety-gated LLM-judge rubric for AI-generated phishing simulations, validated via adversarial-degradation and cross-model agreement testing before production use.
Optimized models for resource efficiency, enabling daily operation across millions of emails at a fraction of the prior compute cost.

2019 — 2022

Elastic

Led Security ML as a function at Elastic — owned the research agenda, scaled the team 3× in 18 months, and shipped ML-backed initiatives like automated alert triage end-to-end from data to production.
Co-authored a passive-aggressive learning method to patch deployed malware-detection models to individual customer environments post-deployment, cutting false positives 23× without retraining or sharing sensitive data.
Credited as a MITRE ATT&CK contributor for sub-technique T1547.009 (Shortcut Modification), a persistence technique covering adversary abuse of startup-folder shortcuts.
Designed ProblemChild, a graph-based framework surfacing anomalous parent-child process relationships from endpoint telemetry to catch living-off-the-land and post-exploit behavior that single-process rules miss.
Built an ML service on top of alert feedback that reduced global false positives by ~40% within 48 hours of model releases.

2015 — 2019

Endgame (acquired by Elastic)

Led a distributed team of data scientists and engineers at Endgame, shipping ML features and data-pipeline enablers while serving as independent monitor across ML projects — tracking progress and reporting to key stakeholders.
Designed and developed Artemis, one of the earliest natural-language agents for security analysts — a conversational interface to EDR event data that drove triage and hunt workflows in place of a structured query language.
Red-teamed production ML/security models (PE-malware classifiers, tree-based detectors) via adversarial example generation and reinforcement-learning-based evasion techniques to surface blind spots ahead of deployment.
Built and open-sourced malware_rl, an RL framework for adversarially evading ML malware classifiers (Ember, MalConv); adopted as reference tooling in subsequent academic research (cited in ICML 2024).

2011 — 2015

Battelle Memorial Institute

Designed and implemented a social-media collection, analysis, and visualization platform.
Developed an authorship-attribution model for source-code attribution.
Built a model demonstrating information diffusion within small online communities and predicting future propagation.

Selected publications

Bertiger, Filar, et al. — Evaluating LLM-Generated Detection Rules in Cybersecurity. CAMLIS 2025. arXiv:2509.16749
Brundage, et al. — The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation. arXiv preprint, 2018 (co-authored with FHI, OpenAI, CSER, EFF, CNAS). arXiv:1802.07228
Anderson, Kharkar, Filar, et al. — Learning to Evade Static PE Machine Learning Malware Models via Reinforcement Learning. arXiv preprint, 2018. arXiv:1801.08917
Filar, et al. — Ask Me Anything: A Conversational Interface to Augment Information Security Workers. USENIX SOUPS WSIW 2017. Workshop paper
Raff, Fleshman, Zak, Anderson, Filar, McLean — Classifying Sequences of Extreme Length with Constant Memory Applied to Malware Detection. AAAI 2021. OJS link

For the full list, see Publications.

Program Committees

CAMLIS — Conference on Applied Machine Learning in Information Security
WoRMA 2026 — 5th Workshop on Rethinking Malware Analysis (co-located with IEEE EuroS&P 2026, Lisbon)
ACM AISec — ACM Workshop on Artificial Intelligence and Security