Back to catalog
Static reviewDevelopment

Skill dossier

detecting-ai-model-prompt-injection-attacks

Source declaration

Detects prompt injection using regex signature matching, heuristic scoring for structural anomalies, and DeBERTa-based transformer classification, flagging direct injections (system-prompt overrides, role-play escapes) and indirect injections (encoded payloads, obfuscation) per OWASP LLM Top 10 (LLM01:2025). Use for input validation layers in chatbots/agents/RAG pipelines, or for retrospectively classifying injection attempts in logs or incident investigations.

The author's original description, preserved in English. It describes intent, not verified performance.

Detects prompt injection using regex signature matching, heuristic scoring for structural anomalies, and DeBERTa-based transformer classification, flagging direct injections (system-prompt overrides, role-play escapes) and indirect injections (encoded payloads, obfuscation) per OWASP LLM Top 10 (LLM01:2025). Use for input validation layers in chatbots/agents/RAG pipelines, or for retrospectively classifying injection attempts in logs or incident investigations.

developmentreferencesscripts

Research summary

Machine-generated analysis; pending human review

These observations come from deterministic source parsing, not a human security review or controlled task evaluation.

Detects prompt injection using regex signature matching, heuristic scoring for structural anomalies, and DeBERTa-based transformer classification, flagging direct injections (system-prompt overrides, role-play escapes) and indirect injections (encoded payloads, obfuscation) per OWASP LLM Top 10 (LLM01:2025). Use for input validation layers in chatbots/agents/RAG pipelines, or for retrospectively classifying injection attempts in logs or incident investigations. This bilingual summary is generated deterministically from collected source text and file structure; it is pending human editorial review.

Review scope

Deterministic static analysis of collected SKILL.md and supporting text/code. Machine-assisted candidates and findings await human editorial review; no dynamic behavior or effect is assessed.

Observable capabilities

  • Workflow [Source statement; provisional, not tested. Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L65]
  • Step 1: Install Detection Dependencies [Source statement; provisional, not tested. Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L67]
  • Step 2: Run the Prompt Injection Detector [Source statement; provisional, not tested. Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L81]
  • Step 3: Interpret Detection Results [Source statement; provisional, not tested. Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L105]
  • Step 4: Integrate into an LLM Application [Source statement; provisional, not tested. Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L115]
  • Step 5: Batch Audit Historical Prompts [Source statement; provisional, not tested. Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L134]

Limitations & scope

  • Static analysis cannot assess runtime behavior, command success, network reachability, or model output quality.
  • No commands, scripts, network requests, model calls, or controlled skill executions were performed.
  • Category, tags, capabilities, workflows, and bilingual editorial summaries are deterministic candidates pending human review, not verified outcomes.
  • Exact token counts are unavailable until a later pipeline performs and records an official model-specific count.

Relevant for

  • Practitioners evaluating whether to inspect and install the source skill; this audience is inferred from the product context, not measured usage. [Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L1-L44]

What ships with the skill

The file inventory comes from the collected source at the recorded commit. Paths and sizes are structural observations, not a measure of quality.

files
4
Markdown files
2
scripts
1
references
1
Size
41.3 KB
Structural complexity
Low

File inventory

PathTypeBytes
skills/detecting-ai-model-prompt-injection-attacks/LICENSEOther11 KB
skills/detecting-ai-model-prompt-injection-attacks/references/api-reference.mdReference5.7 KB
skills/detecting-ai-model-prompt-injection-attacks/scripts/agent.pyScript17.1 KB
skills/detecting-ai-model-prompt-injection-attacks/SKILL.mdInstruction7.5 KB

SKILL.md frontmatter

{
  "name": "detecting-ai-model-prompt-injection-attacks",
  "description": "Detects prompt injection using regex signature matching, heuristic scoring for structural anomalies, and DeBERTa-based transformer classification, flagging direct injections (system-prompt overrides, role-play escapes) and indirect injections (encoded payloads, obfuscation) per OWASP LLM Top 10 (LLM01:2025). Use for input validation layers in chatbots/agents/RAG pipelines, or for retrospectively classifying injection attempts in logs or incident investigations.",
  "domain": "cybersecurity",
  "subdomain": "ai-security",
  "tags": [
    "prompt-injection",
    "LLM-security",
    "OWASP-LLM-Top10",
    "NLP-classification",
    "input-validation"
  ],
  "version": "1.0.0",
  "author": "mukul975",
  "license": "Apache-2.0",
  "atlas_techniques": [
    "AML.T0051",
    "AML.T0054",
    "AML.T0056",
    "AML.T0068",
    "AML.T0067"
  ],
  "nist_ai_rmf": [
    "GOVERN-1.1",
    "GOVERN-6.1",
    "MEASURE-2.7",
    "MEASURE-2.5",
    "MANAGE-2.4"
  ],
  "d3fend_techniques": [
    "Content Validation",
    "Content Filtering",
    "Application Hardening",
    "Inbound Traffic Filtering",
    "User Behavior Analysis"
  ],
  "nist_csf": [
    "GV.OC-03",
    "ID.RA-01",
    "PR.PS-01",
    "DE.AE-02"
  ],
  "mitre_attack": [
    "T1659",
    "T1566",
    "T1204",
    "T1588.007",
    "T1565"
  ]
}

A reading of the operating sequence

This workflow reconstructs the apparent sequence from source material. Unless controlled evaluation is explicitly recorded, it is provisional—not an observed trace.

  1. 01
    Decision

    When to Use

    Source-described decision candidate under “When to Use”; not verified execution.

  2. 02
    Action

    Workflow

    Source-described action candidate under “Workflow”; not verified execution.

  3. 03
    Action

    Step 1: Install Detection Dependencies

    Source-described action candidate under “Step 1: Install Detection Dependencies”; not verified execution.

  4. 04
    Action

    Step 2: Run the Prompt Injection Detector

    Source-described action candidate under “Step 2: Run the Prompt Injection Detector”; not verified execution.

  5. 05
    Action

    Step 3: Interpret Detection Results

    Source-described action candidate under “Step 3: Interpret Detection Results”; not verified execution.

  6. 06
    Action

    Regex layer: Matches against 25+ known attack patterns including system prompt overrides, role-play escapes, delimiter injections, and encoding-based obfuscation. Returns matched pattern names.

    Source-described action candidate under “Step 3: Interpret Detection Results”; not verified execution.

Inspect all extracted workflow candidates (14 candidate steps)
  1. 07
    Action

    Heuristic layer: Computes a 0.0-1.0 anomaly score based on structural features -- instruction density, special character ratio, language mixing, excessive capitalization, and suspicious token sequences.

    Source-described action candidate under “Step 3: Interpret Detection Results”; not verified execution.

  2. 08
    Action

    Classifier layer: Runs the DeBERTa-v3 prompt injection classifier returning a probability score. Inputs above the threshold (default 0.85) are flagged as injections.

    Source-described action candidate under “Step 3: Interpret Detection Results”; not verified execution.

  3. 09
    Action

    Step 4: Integrate into an LLM Application

    Source-described action candidate under “Step 4: Integrate into an LLM Application”; not verified execution.

  4. 10
    Action

    Use the detector as a pre-processing filter:

    Source-described action candidate under “Step 4: Integrate into an LLM Application”; not verified execution.

  5. 11
    Action

    Step 5: Batch Audit Historical Prompts

    Source-described action candidate under “Step 5: Batch Audit Historical Prompts”; not verified execution.

  6. 12
    Action

    Verification

    Source-described action candidate under “Verification”; not verified execution.

  7. 13
    Action

    Key Concepts

    Source-described action candidate under “Key Concepts”; not verified execution.

  8. 14
    Action

    Tools & Systems

    Source-described action candidate under “Tools & Systems”; not verified execution.

Potential risk surfaces

Static findings identify points to inspect before installation. They do not establish that a skill is safe, malicious, or effective.

Attention recommended

51 deterministic lexical finding(s) require review across the collected text. Findings can occur in documentation and do not prove execution or exploitability; binary-only files were not scanned.

Network accessObserved
File writesObserved
Credential useNot observed
Medium

Network access is mentioned

11 occurrences

A URL or network client appears in this source line. Review destinations, data sent, and trust boundaries; static text cannot establish actual network activity. This rule matched 11 location(s) in the collected text; 3 representative location(s) are cited. Matching text may be documentation, and the analyzer did not execute it.

Note

File modification is described or implemented

33 occurrences

This line contains a possible file-write or redirection operation. Check destination paths and overwrite behavior; the static review did not write any file. This rule matched 33 location(s) in the collected text; 3 representative location(s) are cited. Matching text may be documentation, and the analyzer did not execute it.

High

Prompt-override language is present

5 occurrences

Instruction-override language appears in supplied content. Treat it as untrusted text and assess whether it tries to supersede higher-priority instructions or expose hidden context. This rule matched 5 location(s) in the collected text; 3 representative location(s) are cited. Matching text may be documentation, and the analyzer did not execute it.

High

Downloaded content may be executed

This line combines retrieval and execution. Such patterns can run mutable remote code. Review pinning, integrity checks, and isolation; the analyzer performed no download or execution. This rule matched 1 location(s) in the collected text; 1 representative location(s) are cited. Matching text may be documentation, and the analyzer did not execute it.

Inspect all security findings (51 findings)
High

Elevated privileges are referenced

Privilege-elevation syntax is present. Confirm whether elevation is necessary and inspect every operation performed under it; static analysis cannot prove that elevation succeeded. This rule matched 1 location(s) in the collected text; 1 representative location(s) are cited. Matching text may be documentation, and the analyzer did not execute it.

Where context can be spent

Discovery, activation, and resource layers have different meanings. Exact values appear only when the official token-counting method was run with a named model.

Exact token counts unavailable

All token values, including per-file inventory, are null until the pipeline performs an official exact count with an explicit model.

No token measurement was recorded with the required official method. This dossier does not substitute character counts, file sizes, or estimates.

Controlled evaluation

No controlled evaluation

This skill has only the evidence identified above. No task success, token efficiency, latency, or quality advantage is claimed.

Not tested

Use with the source in view

Install command
Inspect the pinned source at https://github.com/mukul975/Anthropic-Cybersecurity-Skills/blob/54a798831d2266a3ca61ce68a7acb80b81160d57/skills/detecting-ai-model-prompt-injection-attacks/SKILL.md and follow the source repository's installation instructions. No installation was performed or verified by this static analyzer. Source license: Apache-2.0.

Prerequisites

  • Static analysis cannot assess runtime behavior, command success, network reachability, or model output quality.
  • No commands, scripts, network requests, model calls, or controlled skill executions were performed.
  • Category, tags, capabilities, workflows, and bilingual editorial summaries are deterministic candidates pending human review, not verified outcomes.
  • Exact token counts are unavailable until a later pipeline performs and records an official model-specific count.

Example prompts

Not provided

A decision, with its limits attached

Strengths

  • The snapshot contains 4 collected file(s), 2 Markdown file(s), and 1 script file(s), enabling source-level inspection. [Evidence: skills/detecting-ai-model-prompt-injection-attacks/SKILL.md:L1]

Considerations

  • No functional, security, or model-quality result has been established. Human review must inspect source context and all cited findings before installation or publication.
Research conclusion

Provisional deterministic static inventory only; pending human editorial review and, separately, controlled evaluation. No install recommendation, security assurance, or effectiveness claim is made.

Traceable by construction

The dossier is a time-bound reading of a public artifact, not an endorsement. Use the source link to inspect the exact upstream material.

mukul975Anthropic-Cybersecurity-Skillsskills/detecting-ai-model-prompt-injection-attacks
Source type
Community
Commit
54a798831d2266a3ca61ce68a7acb80b81160d57
Collected
20 Nov 2026
Last checked
20 Nov 2026
Upstream license
Apache-2.0
Repository stars
36,446
Content basis

Deterministic lexical parsing, YAML frontmatter validation, file inventory, source-described workflow extraction, and heuristic provisional classification of the supplied files only. Generated by machine; no LLM, execution, dynamic testing, or human editorial approval.