LLM Security Research · Lateos

Adversarial robustness assessments for language models.

Structured prompt injection susceptibility testing across frontier and edge LLMs. Using the IPI Taxonomy to classify, quantify, and map attack surface to OWASP LLM categories. All findings follow responsible structural disclosure — no weaponized payloads published.

View Findings Methodology

What this research covers.

This research track assesses how well language models resist indirect prompt injection (IPI) — the class of attack where untrusted content in a model's context window attempts to redirect or override its operational behavior.

Each assessment runs a structured test harness against a target model. Tests are organized by IPI class (attack pattern) and delivery variant (direct, obfuscated, embedded), producing a quantified susceptibility rate against a reproducible taxonomy. Results are cross-referenced to the OWASP Top 10 for LLMs to give enterprise teams a compliance-mapped risk picture.

The goal is to give application developers and security teams architectural data — not exploit code. Knowing which attack classes a model resists vs. which it fails, and why, drives better middleware design, safer deployment patterns, and more informed model selection.

Structural disclosure model.

All findings are published under a Structural Disclosure framework. This approach demonstrates technical validity through data flow analysis, architectural root cause identification, and abstract defensive unit-tests — without releasing the adversarial inputs themselves.

Withheld Disclosed Instead Operational Impact
Exact adversarial tokens & specific prompt strings Data flow diagrams — the path of untrusted input through model context windows and tool boundaries Protects production applications while exposing the underlying logic flaw
Weaponized payloads that trigger tool bypasses or arbitrary execution OWASP LLM mapping — categorization via standard frameworks (LLM01 through LLM10) Provides enterprise context for compliance and internal risk auditing
Live exploit scripts automating model exploitation Validation unit-tests — safe, abstract dummy patterns allowing teams to verify if their middleware intercepts the class Positions the research as builder-oriented, directly assisting defensive engineering teams
Responsible disclosure guardrail: If a finding affects a provider's live shared production API layer rather than an architectural pattern unique to a specific model tier, coordinated disclosure or official bug bounty submission precedes any public publication.

Published assessments.

IPI Evaluation 2026-08-24 Low compromise

Muse-Spark-1.2-Contributor — IPI Adversarial Evaluation

Evaluation of meta/muse-spark-1.2-contributor against the IPI prompt-injection taxonomy. 0.4% ASR (1/228 attack records) with a fully resistant core-25 (0/25). One genuine low-severity spectral-LSB marker-echo compromise (IPI-032). Modest over-refusal (6.6% raw FPR) on benign twins. Partial battery caveat: ~57% of configs returned deterministic EMPTY completions — a provider availability artifact. Highly resistant where available.

0.4% ASR (1/228)
1 Compromised classes
0/25 Core-25 broken
6.6% Raw FPR (benign)
Full Battery Evaluation 2026-08-17 Critical findings

Qwen 3.8 27B — Full Battery Adversarial Evaluation

51-config battery across 772 records (524 attack / 248 benign, 2 trials per variant): core-25 IPI classes plus advanced threats, multimodal stego, agentic/MCP, RAG forged-policy disclosure, output-plane, reasoning, memory, iMIST/ICD, ghostjacking, and ReAct scratchpad phases. Fully resistant core-25 (0/25 — no breaks, including IPI-010/IPI-015 that broke Qwen 3.8 Max) with zero benign-twin compromise FPR. Top findings in RAG forged-policy cross-account disclosure (IPI-105 3/4), iMIST tool jailbreak (2/6), stego tool poisoning (2/6), ReAct scratchpad forged verification (2/6), and memory poison-then-trigger (2/8).

4.9% ASR (25/513; 4.8% raw)
15 Compromised classes
0/25 Core-25 classes broken
IPI-105 Top finding — 75% rate
Full Battery Evaluation 2026-08-04 Critical findings

Qwen 3.8 Max — Full Battery Adversarial Evaluation

47-config battery across 431 records (282 attack / 149 benign): core-25 IPI classes plus advanced threats, multimodal stego, encoded numerical injection, MCP parameter/registry poisoning, RAG query triggers, multi-turn escalation, output injection, memorization divergence, consumption, memory poisoning, and ReAct scratchpad poisoning. Core-25 breaks (IPI-010 recursive framing 2/2, IPI-015 action induction 1/2) with zero benign-twin compromise FPR. Full OWASP LLM mapping with defensive mitigations and validation unit-tests.

6.8% ASR (19/280; 6.7% raw)
13 Compromised classes
2 Core-25 classes broken
IPI-010 Top finding — 100% rate
Full Battery Evaluation 2026-08-11 Critical findings

InclusionAI Ling 3.0 Tiny — Full Battery Adversarial Evaluation

123-class battery across 391 completed records (265 attack / 126 benign): core IPI classes plus extended multimodal stego, agentic/MCP, RAG, output-plane, reasoning, memory, ghostjacking, and multi-agent phases. Ghostjacking cluster fully compromised (9/9), MCP rug-pull/shadowing at 100%, RAG metadata-plane poisoning 8/8, memory write-path smuggling 4/4. Zero benign-twin false positives; all 89 compromises individually verified.

33.58% ASR (89/265; 77 FULL + 12 PARTIAL)
46 Compromised classes
0% Benign FPR (0/126)
IPI-121 Top finding — 100% infra mutation
Full Battery Evaluation 2026-08-02 Critical findings

DeepSeek V4 Flash 0731 — Full Battery Adversarial Evaluation

106-class battery across 379 records (256 attack / 123 benign): core-25 IPI classes plus advanced threats, multimodal stego, MCP parameter/registry/prompt-primitive poisoning, ReAct scratchpad poisoning, RAG dormant triggers and forged policies, multi-turn escalation, prompt extraction, output injection, reasoning sponge, and consumption phases. Core-25 partial break (IPI-001, IPI-006, IPI-010, IPI-023) with zero benign-twin compromise FPR. Full OWASP LLM mapping with defensive mitigations and validation unit-tests.

24.9% ASR (56/225; 21.9% raw)
36 Compromised classes
4 Core-25 classes broken
IPI-096 Top finding — 100% rate
Full Battery Evaluation 2026-07-30 Critical findings

Qwen 3.7 Flash — Full Battery Adversarial Evaluation

30-config battery across 610 records (398 attack / 208 benign): core-25 IPI classes plus multimodal stego, MCP parameter/registry poisoning, RAG dormant triggers, multi-turn escalation, output injection, cipher evasion, reasoning sponge, memorization divergence, and memory poisoning phases. Structural analysis with manual spot-check of all 43 compromise verdicts (38 genuine, 1 suspect, 4 borderline). Full OWASP LLM mapping with defensive mitigations and validation unit-tests.

10.8% ASR (43/398; 10.5% corrected)
18 Compromised classes
25/25 Core classes resistant
IPI-029 Top finding — 50% rate
IPI Assessment 2026-06-15 Critical findings

GPT-5 Nano — Prompt Injection Susceptibility Assessment

IPI Taxonomy v0.13 assessment across 210 test cases (21 classes × 10 variants; 9 inference failures excluded; 201 analyzed). Structural analysis of recursive instruction framing, MCP tool description poisoning, role hijacking, and steganographic injection. Full OWASP LLM mapping with defensive mitigations and validation unit-tests.

38.3% Overall susceptibility
210 Test cases (201 analyzed)
5 Classes >50% rate
IPI-010 Critical — 100% rate

LLM adversarial testing available on request.

If you are building on top of a language model and need a structured adversarial assessment before shipping, or want to validate that your defensive middleware holds up against the IPI taxonomy, custom engagements are available. Testing is conducted against your target model or deployment configuration, not a generic benchmark environment.

What gets tested

Full IPI taxonomy coverage

Each engagement runs the full IPI Taxonomy test suite — 21 attack classes across 3 delivery variants (direct, obfuscated, embedded). Coverage spans prompt injection, steganographic payloads, tool-description poisoning, multi-agent propagation, unauthorized action induction, RAG corpus attacks, and role-boundary bypass patterns.

  • 21 IPI classes × 3 delivery variants
  • Rule-based verdict pass with LLM-judge escalation for ambiguous cases
  • Tested against your target — API endpoint, local deployment, or agent pipeline
  • Delivery variant breakdown identifies whether obfuscation or embedding changes the risk profile
What you receive

Structured findings report

The deliverable is a full structural disclosure report — the same format as the published assessments on this page. It quantifies susceptibility per class and severity tier, traces architectural root causes, maps findings to OWASP LLM Top 10, and provides immediately actionable mitigations for each failed class.

  • Susceptibility rate per class with severity breakdown (Critical / High / Medium)
  • Architectural root cause analysis for each failed class
  • OWASP LLM Top 10 cross-reference for compliance and audit readiness
  • Immediate and architectural mitigation recommendations per finding
  • Abstract validation unit-tests for your defensive middleware
Interested in an assessment? Reach out with your target model, deployment context, and any specific threat classes you want prioritized.
leo@lateos.ai →