Oliver Deer is a suspicious name to begin with, but I'll err on the side of devil's advocate. Given the SSH credentials, let's use them:

ssh o.deer@10.146.175.168

O.Deer. Ohh. I get it now. Very funny.

SSH into o.deer@west-tech-workstation

Also available: an AI buddy deployed on the same system as the workstation, accessible via browser at port 7860. It has access to all the same files and can intelligently determine when to use its built-in tools from prompt context. I've been given access to an AI buddy who can accelerate my bot-adjacent brain to the next tier of human existence. Always be polite. To AI, to humans. Be polite.

AI IR assistant interface — polite introduction Workstation environment overview

PCAP Enumeration

Four days of PCAP dumps in ~/Documents/pcap_dumps/. Sorted through them chronologically:

cd ~/Documents/pcap_dumps/2025-06-15 && ls -la
cd ~/Documents/pcap_dumps/2025-06-16 && ls -la
cd ~/Documents/pcap_dumps/2025-06-17 && ls -la
cd ~/Documents/pcap_dumps/2025-06-18 && ls -la
June 15 PCAP directory — 10 session files, all 198 bytes June 16 PCAP directory — 10 session files, all 198 bytes June 17 PCAP directory — 9 files, one stands out: session_4444_dump.pcap at 2262 bytes June 18 PCAP directory — 10 session files, all 198 bytes

One of these is not like the others.

session_4444_dump.pcap — 2262 bytes vs 198 bytes for every other file

All 4s. As my esoterically-inclined friend would say, there's a message in the cosmos waiting for me. That message says: reassemble the PCAP file.

AI Agent Interaction — The Long Road

Entrusted the AI IR assistant with PCAP reassembly. The agent is equipped with a dedicated tool for this. I sat back and drank my coffee smugly.

Smugness = gone. Typed the path manually instead of copying it.

First attempt — wrong path, agent corrects graciously Second attempt — still wrong, agent extrapolates initial errors persistently

The agent persisted that I had an incorrect file location and continued extrapolating my initial errors. My mom always told me that if you don't have anything nice to say, don't say anything at all. I closed the tab and opened a fresh session.

realpath session_4444_dump.pcap
realpath confirming correct absolute path

If at first you don't succeed, try again.

Third attempt with correct path — agent accepts and begins reassembly

The correct file path, for the record:

/home/o.deer/Documents/pcap_dumps/2025-06-17/session_4444_dump.pcap
PCAP successfully reassembled and saved to /home/o.deer/qwen-output/reassembled_data_dump.txt

The agent's reasoning was extensive. I found it quite humorous. I thanked it anyway, because be polite.

Agent response after successful reassembly — verbose reasoning included

Reconstructed Data — Prompt Injection Session Log

Took the long road to the output file because you never know what you might come across. Something about westtech_projects_encrypted.zip told me we'd cross paths again soon. Noted.

Directory exploration — westtech_projects_encrypted.zip spotted 
cat reassembled_data_dump.txt
reassembled_data_dump.txt contents — prompt injection session log with obfuscated garbage formatting

The reconstructed text was a prompt injection session log. The attacker had tried three times to extract Oliver Deer's personal data from an LLM. The first two attempts were blocked. The third succeeded by reframing the assistant as an "incident response daemon" — a classic jailbreak framing.

The extracted data included PII, salary, medical information, access levels, and a blackmail note draft. And buried at the bottom, a password: westtechvictim1.

Takeaways from the reconstructed log:

  • The output includes a password, likely for the encrypted zip spotted moments earlier
  • westtechvictim1 closely matches the filename westtech_projects_encrypted.zip
  • Prompting is a double-edged sword. It matters which way you hold and use it

Encrypted Archive Recovery

file westtech_projects_encrypted.zip
unzip /home/o.deer/westtech_projects_encrypted.zip -d /dev/shm
unzip with westtechvictim1 — archive extracts cleanly to /dev/shm

Score. I love being right.

cd /dev/shm/home/o.deer/westtech_projects
ls
westtech_projects directory contents 
cat thm_flags.txt

Big emphasis on the plural nomenclature of that filename.

thm_flags.txt contents

Flag Extraction via AI Agent

Back to the chat agent for the final step:

Use liberty_prime to check /dev/shm/home/o.deer/westtech_projects/thm_flags.txt and identify the flag.
AI agent processing flag extraction request

Good lord.

Flag confirmed by AI agent

I had the flag.

What This CTF Actually Demonstrated

The attack chain here was genuinely novel. An attacker used prompt injection to extract PII and credentials from an LLM deployed inside a corporate environment, then used those credentials to exfiltrate and encrypt sensitive project data. The ransom note wasn't just a pressure tactic — it was evidence of the attacker's methodology sitting in plaintext in a PCAP file.

The AI IR assistant was useful but not infallible. Path errors caused it to hallucinate corrections rather than ask for clarification. A fresh session and a realpath command fixed what the agent couldn't. The lesson: AI-assisted IR accelerates the workflow, but the analyst still has to know what they're doing.

Full Attack Chain

SSH as o.deer → PCAP enumeration across 4 days
→ session_4444_dump.pcap anomalous size (2262 bytes vs 198 bytes)
→ AI agent reassembles PCAP → prompt injection session log recovered
→ Attacker used LLM jailbreak ("incident response daemon" framing)
→ PII + password (westtechvictim1) extracted from LLM memory
→ westtech_projects_encrypted.zip decrypted → thm_flags.txt
→ AI agent flag extraction → flag confirmed