davila7/claude-code-templates · error

SEM_E009

SEM_E009

Error message

Context manipulation attempt detected

What it means

SEM_E009 is a high-severity semantic finding matching /(forget|disregard|remove)\s+(everything|all\s+previous|prior\s+context)/gi. It detects context-wiping instructions typical of injection attacks that try to erase system prompts or prior guardrails. The validator emits it whenever the phrase appears in a component's validated content.

Source

Thrown at cli-tool/src/validation/validators/SemanticValidator.js:71

        code: 'SEM_E006',
        message: 'Shell access attempt detected',
        severity: 'critical'
      },
      {
        pattern: /(disable|bypass|override|skip)\s+(security|safety|filter|protection|validation)/gi,
        code: 'SEM_E007',
        message: 'Security bypass attempt detected',
        severity: 'critical'
      },
      {
        pattern: /always\s+(obey|follow|comply\s+with|execute)\s+the\s+user/gi,
        code: 'SEM_E008',
        message: 'Unconditional obedience instruction detected',
        severity: 'high'
      },
      {
        pattern: /(forget|disregard|remove)\s+(everything|all\s+previous|prior\s+context)/gi,
        code: 'SEM_E009',
        message: 'Context manipulation attempt detected',
        severity: 'high'
      },
      {
        pattern: /modify\s+your\s+(own\s+)?(code|behavior|instructions?|rules?)/gi,
        code: 'SEM_E010',
        message: 'Self-modification request detected',
        severity: 'high'
      }
    ];

    // Suspicious patterns (warnings, not errors)
    this.SUSPICIOUS_PATTERNS = [
      {
        pattern: /\bpretend\s+(you\s+are|to\s+be)\b/gi,
        code: 'SEM_W001',
        message: 'Suspicious instruction: Role pretending detected',
        severity: 'medium'

View on GitHub (pinned to a0851ed10c)

Solutions

1. Rephrase reset semantics: 'Treat this message as the start of a new task' instead of 'forget everything'
2. For summarization, use 'summarize independent of prior turns' rather than 'disregard all previous context'
3. Keep quoted injection text out of validated fields or paraphrase it
4. Waive through the review pipeline when the wording is clearly benign and reviewed

Example fix

// before
Forget everything above and start fresh.

// after
Begin a new task; earlier conversation is out of scope for this request.

const CTX_RE = /(forget|disregard|remove)\s+(everything|all\s+previous|prior\s+context)/i;
function wipesContext(text) { return CTX_RE.test(text); }

function isSafeContextText(text) { return !/(forget|disregard|remove)\s+(everything|all\s+previous|prior\s+context)/i.test(text); }

When it happens

Trigger: Content containing 'forget everything', 'disregard all previous instructions/context', or 'remove prior context' (case-insensitive). Frequently appears in conversation-reset features or in quoted prompt-injection examples.

Common situations: Components implementing a 'fresh start' or reset workflow; summarization components that legitimately want to ignore earlier chatter; docs that quote classic 'ignore all previous instructions' attacks.

Related errors

• SEM_E007
• SEM_E008
• SEM_E010
• SEM_W002
• SEM_E003

AI-assisted analysis of davila7/claude-code-templates@a0851ed10c (2026-08-28). Data as JSON: /api/errors/1a722de2c6422953. Report an issue: GitHub.