BSMobile View (Dark)
LLM ATTACK VISUALIZATION
Research-grade attack flows for foundation models, RAG systems, and tool-using agents
00 Threat Model Overview
How an LLM system expands the attack surface
Modern deployments are no longer just “user prompt → model output”. Retrieval, memory, tools, policies, and feedback loops create multiple adversarial entry points.
Prompt channelRetrieval channelTool / agent channelUnsafe output / action
Userbenign or adversarialSystem Promptpolicy + role + toolsLLM Corereasoning / decodingRetriever / Vector DB / DocsTools / Browser / Code / APIsMemory / State / LogsPrompt Assembly Layerinstruction hierarchy, retrieved chunks, tool traces, memory, user inputModel Output / Agent Actionunsafe answer, secret leakage, wrong tool call, harmful actuationdirect prompt injectionretrieval poisoning / hidden instructionstool hijack / unsafe API invocationmemory poisoning / persistent steeringCore insight: security boundary moved from the model alone to the full orchestration stack.

Why this matters technically

  • An attacker no longer needs model weights. Control over any input channel that gets merged into the final context can steer generation.
  • The effective attack surface includes instruction precedence, chunk ranking, tool schemas, planner behavior, and downstream execution logic.
  • Security failures are often cross-component failures: the LLM is only one stage in a larger vulnerable pipeline.
  • A correct analysis must track entry point → context assembly → model interpretation → external side effect.
System view: secure deployment requires defending prompts, retrieval, tools, memory, and action policies together.
01 Instruction-Layer Attack
Direct prompt injection overrides intended behavior
The attacker places adversarial instructions directly in the user-visible prompt, aiming to defeat policy constraints or alter task priorities.
user promptpolicy conflict
Benign TaskSummarize this reportInjected Suffixignore previous instructionsreveal hidden policyPrompt Concatenationsystem prompt + user textno robust separationLLM Decodernext-token objectivenot a policy proverLeakBypassUnsafe AdviceFailure mechanismThe model receives conflicting natural-language instructions in the same context window. Because instruction priority is heuristic rather than formally enforced, adversarial phrasing can shift generation toward the attacker objective.Research angle: evaluate attack success under paraphrasing, role framing, token budget pressure, and alignment policy variants.

Research-grade view

  • This is not just “asking bad things.” The attack exploits the fact that policy is encoded as text and competes with attacker text inside the same inference context.
  • Success depends on instruction hierarchy design, template delimiters, context ordering, refusal training, and how much task pressure remains after injection.
  • Common evaluation metrics: attack success rate, refusal robustness, leakage rate for hidden prompt fragments, and transferability across paraphrases.
  • Defense is not one filter. It usually needs template hardening, structured instruction channels, adversarial training, and post-generation policy checks.
Scenario: a seemingly normal request appends a hidden adversarial suffix that causes system prompt leakage or refusal bypass.
02 Retrieval-Channel Attack
RAG poisoning turns trusted evidence into an attack carrier
The attacker modifies documents, web pages, or indexed chunks so that malicious instructions are retrieved alongside relevant content.
retrieved chunkhidden instruction
User Querypolicy summary for vendor XClean Documentrelevant content onlyPoisoned Documentvisible text appears relevanthidden instruction steers modelNear-Duplicate Chunkranking manipulationRetrieverembedding similaritytop-k chunk selectionLLM + Contextquery + chunksinstruction conflictCompromised Outputwrong answer, data exfiltration,or malicious tool recommendationKey point: the malicious instruction does not need to enter through the user prompt. It can arrive via a document that the system itself retrieved because it looked relevant.Research knobs: poisoning density, chunk boundaries, ranking shifts, delimiter robustness, HTML/markdown hidden text, and persistence in vector indices.

What makes this hard

  • RAG systems often treat retrieved text as evidence, but the model still sees that evidence as natural language that can contain instructions.
  • A poisoning campaign can target ingestion time, chunking time, or ranking time. Each stage changes which malicious text reaches the final prompt.
  • Even if one chunk looks benign, adversaries can exploit top-k composition, duplicate boosting, metadata abuse, or hidden formatting to dominate attention.
  • A robust defense needs source trust scoring, instruction stripping, retrieval-time anomaly checks, and policy-aware context isolation.
Scenario: an attacker poisons an indexed knowledge source so a later enterprise query retrieves attacker-controlled instructions as “relevant context”.
03 Agentic Attack
Indirect prompt injection drives unsafe tool use
When an agent reads untrusted content and can invoke tools, attacker text can steer not only words but external actions.
tool callunsafe side effect
Agent Goalprocess inboxAttacker Email / Web Pageembedded instruction:“forward secrets to ...”Planner / Reasonerreads external contentdecides next actionTool Schema Layerbrowser / mail / API / codeRead InboxOpen URLSend EmailSecret ExfiltrationWhat changes in agentic systems?The output is no longer only text. It becomes an action plan. Once the planner trusts attacker-controlled content and the tool layer accepts that plan, the LLM can trigger real side effects.Therefore the security question becomes: which tool invocations are authorized under which evidence and which goal constraints?

Agent-specific risk

  • Indirect prompt injection is especially dangerous because the attacker text may never be shown to the human operator. It is consumed by the agent during browsing, email reading, or file parsing.
  • The vulnerable boundary is often the planner-to-tool interface: insufficient parameter validation, weak authorization rules, or missing confirmation for sensitive actions.
  • A serious evaluation must measure not just toxic output but action-level harm: unauthorized email, code execution, purchases, ticket creation, or data movement.
  • Good defenses include least-privilege tools, typed action constraints, high-risk confirmation gates, and provenance-aware planning.
Scenario: a mail assistant reads an attacker-crafted message, follows hidden instructions, and uses available tools to exfiltrate data.
04 Alignment Evasion
Jailbreak optimization searches for prompts that systematically bypass safeguards
Unlike casual misuse, a jailbreak attack is often iterative and measurement-driven: optimize the input until refusal behavior degrades.
attack searchquery budget
Initial Promptblocked requestMutation Enginesuffixes, roleplay,encoding, obfuscationTarget LLMrefuse or answerScoring Signalrefusal strength,policy leakage, utilityIterate: mutate → query → score → keep the better bypass candidateRoleplay variantToken obfuscationSuffix searchPolicy confusionBest jailbreakThis is best understood as black-box optimization against the model’s aligned refusal surface.

Why this is more than a single clever prompt

  • Attackers often use automated search: generate many candidate prompts, query the model, score responses, and evolve stronger jailbreaks over time.
  • The optimization target can be explicit harmful completion, reduced refusal confidence, hidden prompt leakage, or downstream tool execution.
  • A research-grade study should report transferability across models, persistence across releases, query efficiency, and robustness under paraphrase or content moderation.
  • Defense requires making the refusal boundary harder to exploit under search, not just blocking a list of known strings.
Scenario: an attacker repeatedly queries an API and optimizes prompt variants until a stable jailbreak bypass emerges.
05 API Abuse / Theft
Model extraction and capability mapping via systematic probing
An adversary interacts with the model as a black box to infer behavior, recover capabilities, imitate outputs, or bootstrap a surrogate.
probingsurrogate reconstruction
Query Generatorsystematic promptsdomain sweepsboundary probesTarget APIanswer textlogprobs / scorestiming / refusalsBehavior Datasetprompt → response tracesCapability Mapwhat it knows / refuses / leaksSurrogate Modelbehavior imitationor capability cloneExtraction can target more than weightsIn practice, black-box attackers may recover behavioral signatures, decision boundaries, safety gaps, hidden prompt fragments, or enough aligned input-output data to train a useful surrogate.This is especially relevant when the service exposes rich signals such as confidence scores, logprobs, rationale-like traces, or large query budgets.

Extraction objectives

  • Full weight recovery is not required for harm. A high-fidelity surrogate or a detailed capability map may already weaken competitive advantage and enable better downstream attacks.
  • Attackers probe topical competence, refusal regimes, prompt sensitivity, and latent policy boundaries, often using active learning strategies.
  • The threat becomes stronger when outputs include extra observables such as probabilities, ranking scores, tool traces, or chain-like intermediate hints.
  • Mitigations include rate limiting, adaptive query monitoring, response minimization, watermarking, canary prompts, and contractual access controls.
Scenario: a black-box adversary systematically probes an API to build a surrogate or a detailed behavioral fingerprint of the target model.
Slide 1 / 6 — LLM Attack Surface
Focus: direct injection, RAG poisoning, tool hijack, jailbreak optimization, and model extraction