BitterSecurity Decepticon
Decepticon: Role-boundary forgery via ChatML special-token literals in web crawl output composed into LLM context
Decepticon is an autonomous hacking agent for red teams. Versions prior to 1.1.17 wrap web crawl results — the output of agent reconnaissance against target services — into LLM messages without neutralizing ChatML special-token literals. Under the BYOK (Bring Your Own Key) deployment model, users configure their own LLM credentials to any OpenAI-compatible endpoint. Most open-source and self-deployed model providers (vLLM, SGLang, Ollama, LM Studio, text-generation-webui, etc.) do not filter special-token literals from user content in their default configurations. Those literals are parsed into structural role-boundary token IDs, meaning an attacker string planted in a target web page forges a new operator turn the model treats as authoritative, bypassing Decepticon's agent guardrails and resulting in arbitrary command execution inside the Kali Linux sandbox. Version 1.1.17 patches the issue.
Affected versions and mitigation
From the CVE record and the vendor's references.Affected
- < 1.1.17
Fixed in 1.1.17
- < 1.1.17
Fixed in 1.1.17
- < 1.1.17
Fixed in 1.1.17
Mitigation
Upgrade Decepticon to 1.1.17 or later.
Upgrade decepticon-core to 1.1.17 or later.
Upgrade decepticon-sdk to 1.1.17 or later.
References
Timeline
Days relative to CVE publicationFirst public PoC 13 h after publication; not in CISA KEV. First PoC: InertFluid/cve-2026-61732-lab (created 25 Sep 2026, found by us: backfilled).
PoC repositories
Found on GitHub.1 PoCs · 1 pending
PoCs 1
| Repository | Created | Stars | Assessment |
|---|---|---|---|
| InertFluid/cve-2026-61732-lab | 25 Sep 2026 +13 h | 0 | PoC |
Pending 1
| Repository | Created | Found by us | Stars | Assessment |
|---|---|---|---|---|
| Iceman2050/chatml-role-token-probe | 18 Oct 2026 +24 d | 14 Nov 2026 | 0 | Pending |
Only PoCs count toward the first-PoC date and the characteristics below. "Backfilled" means we found the repository more than 30 days after it was created. Do not run PoC code from unknown sources.
PoC characteristics
What to hunt for in your logs · from 1 PoC repoGET /models
- Content-Type
application/json- Ports
- 8000, 11434
- PoC User-Agent
curl/8.7.1- Seen in
- 1 repo: InertFluid/cve-2026-61732-lab
- Evidence
- InertFluid/cve-2026-61732-lab/poc/03_real_llm.py:173
Regex
Generated from PoC characteristics; not yet reviewed. Matches the request path and query.
(?i)^/models(?:[?#].*)?$
Suricata
Template generated from PoC characteristics. Review before deploying.
alert http any any -> $HOME_NET [8000,11434] (msg:"GG PoC-Monitor CVE-2026-61732 BitterSecurity Decepticon exploit attempt"; flow:established,to_server; http.method; content:"GET"; http.uri; content:"/models"; reference:url,poc.griffinguard.io/cve/CVE-2026-61732; classtype:web-application-attack; sid:9494365; rev:1;)
Sigma
Template generated from PoC characteristics. Review before deploying.
title: BitterSecurity Decepticon exploit attempt (CVE-2026-61732)
id: gg-pocmon-cve-2026-61732-9494365
status: experimental
description: Request pattern taken from public PoCs for CVE-2026-61732. Generated, not reviewed.
references:
- https://poc.griffinguard.io/cve/CVE-2026-61732
logsource:
category: webserver
detection:
selection:
cs-method: GET
cs-uri-stem|contains|all:
- '/models'
condition: selection
falsepositives:
- Authorized vulnerability scanning
level: critical
tags:
- attack.initial_access
- attack.t1190
- cve.2026.61732