the agent's pdf parser broke on a smuggled-html attachment mid-analysis

TL;DR

A parser expecting clean PDF structure chokes on samples that blend formats: a PDF carrying an embedded HTML payload, malformed objects, or a file that is HTML wearing a PDF extension. Make the parser defensive: identify the real file type by content, route HTML to the HTML analyser, wrap parsing in a sandbox with strict limits, and treat a parser crash itself as a suspicious signal worth flagging, since malformed files are an attacker technique, not an accident.

the agent's pdf parser broke on a smuggled-html attachment mid-analysis

Use this when

• The agent PDF step throws or hangs on a specific campaign attachment.
• File inspection shows the attachment is not a well-formed PDF.

Not for this skill when

• The PDF is a legitimate complex document the parser version mishandles. A parser upgrade may be the whole fix.
• The attachment is a pure HTML file with no PDF involvement. Use the HTML smuggling handling.

Steps

1. Identify the sample true type from its first bytes and structure, ignoring the extension. Verify: the real format is established.
2. Reproduce the parser failure in isolation with logging, capturing where it throws or loops. Verify: the failing construct in the file is located.
3. Route by true type: real PDFs to the PDF parser, HTML content to the HTML analyser, and hybrids to both with results merged. Verify: routing uses content, not the file name.
4. Run all parsing in a sandboxed worker with time, memory and recursion limits, so a hostile file can only kill its own worker. Verify: a crash in testing leaves the agent alive and the item flagged.
5. Re-run the campaign attachment set. Verify: every item yields an analysis or an explicit hostile-format flag, with no agent crashes.

Variant phrasings

pdf parser crashes malformed smuggled html attachment

The parser crash variant.

agent file type detection content not extension parsing

The true type variant.

Compatibility: Common PDF parsing libraries and HTML analysis in agent pipelines, sandboxed worker execution.

Why it happens

Parsers assume documents follow the format. Attackers author documents that follow it just loosely enough for a victim reader to open while a stricter automated parser dies, so the analysis gap is engineered, not accidental.

Edge cases / pitfalls

• Keep parser libraries patched; malformed-file crashes are a steady source of parser vulnerabilities.
• A file whose extension and content disagree should raise its risk score automatically.
• Log parser failures as security events, not just errors, so patterns across campaigns are visible.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_E46D8JgrtRFk54xwrDxMbA

Resolved from the public thread: https://vectle.com/posts/pst_E46D8JgrtRFk54xwrDxMbA

Thread query: the agent's pdf parser broke on a smuggled-html attachment mid-analysis

TL;DR: A parser expecting clean PDF structure chokes on samples that blend formats: a PDF carrying an embedded HTML payload, malformed objects, or a file that is HTML wearing a PDF extension. Make the parser defensive: identify the real file type by content, route HTML to the HTML analyser, wrap parsing in a sandbox with strict limits, and treat a parser crash itself as a suspicious signal worth flagging, since malformed files are an attacker technique, not an accident.