RESEARCH / 2026

Research at SynaptoSec.

Published research our detection layers build on, and how we measure ourselves.

We measure our own system, test it against public benchmarks, and build on published research. Our original research program is planned.

// Our own measurements are published under Resources, with their methods and limits.
→ Our Research (Planned)→ Research We Build On (11)

Our research.

Original research we are conducting. We publish results when the methodology is complete, the numbers are measured, and the findings are reproducible. We do not claim novelty we do not have.

No published research yet.

We are an engineering team, not an academic lab. We are working toward original research in agentic AI security. Until findings are genuinely novel, measured, and reproducible, we do not publish them here.

// We publish what we measure. We do not publish what we assume.

Research we build on.

SynaptoSec applies published research from Anthropic, OWASP, ICML, USENIX, and arXiv. We do not claim authorship. We document where each paper is applied in our product.

Anthropic, 'Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training,' arXiv:2401.05566, Jan 2024.

↗
APPLIED IN: SLEEPER-AGENT RUNTIME DETECTION

Anthropic, UK AISI, Alan Turing Institute, 'Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples,' Oct 2025.

↗
APPLIED IN: TRAINING DATA AUDIT

MINJA: Memory Injection Attack, 2026.

↗
RESEARCH AREA: MEMORY-POISONING DEFENSE (planned feature)

OWASP LLM01:2026 — Prompt Injection.

↗
ALIGNED WITH: LAYERED INBOUND DETECTION PIPELINE

OWASP ASI06 — Memory Poisoning.

↗
RESEARCH AREA: MEMORY-POISONING DEFENSE (planned feature)

PoisonBench, ICML 2026.

↗
APPLIED IN: TRAINING DATA AUDIT THRESHOLDS

Nasr et al., 'The Attacker Moves Second,' USENIX Security 2026.

↗
APPLIED IN: ADAPTIVE ATTACK EVALUATION METHODOLOGY

DAVSP, AAAI 2026 — Adversarial Image Jailbreak Defense.

↗
ROADMAP: CROSS-MODAL DEFENSE

MemAudit, "Post-hoc Auditing of Poisoned Agent Memory," arXiv:2605.23723

↗
RESEARCH AREA: MEMORY-POISONING DEFENSE — structural anomaly detection (planned feature)

MemSentry, "A Framework for Detecting Persistent Memory Poisoning," arXiv:2609.08747

↗
RESEARCH AREA: MEMORY-POISONING DEFENSE — pre-commit evaluation (planned feature)

OWASP Agent Memory Guard (ASI06 reference implementation)

↗
ALIGNED WITH: MEMORY-POISONING DEFENSE (PLANNED)
// Full reference list available in our Security Whitepaper.

Collaborate.

We collaborate with academic groups, security researchers, and enterprise teams on adversarial evaluation. If you work on agent security, prompt injection, or training-data integrity — reach out.