NVD published 24 Sep 2026
CVE-2026-61732
2 public exploits.
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.
Signals
- KEV, exploited in the wild
- no
- EPSS, 30-day forecast
- 1.5%, 74th percentile (as of 18 Nov 2026)
- CVSS, NVD, CRITICAL
- 10.0
- Vector
- CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H
- Internet exposure
- not stated
Not in CISA's catalogue as of the last daily pull. Absence is not proof of safety.
Timeline 24 Sep 2026 to today
- Disclosed
- First PoC+1 d
1 days from disclosure to the first proof of concept
Indicators none tied yet
No threat feed has tied an indicator to this CVE yet. Checked every six hours.
On the wire 1 dispatch
Public exploit code 2 known · links only
GitHub PoCs 2Metasploit 0Exploit-DB 0
GitHub PoCs, by trust then stars2
- Iceman2050/chatml-role-token-probeTrust ok 1000 stars
Fire ChatML/Mistral/Llama control-token literals at your self-hosted agent stack and prove role-forgery before an attacker does (CVE-2026-61732).
created 18 Oct 2026
- InertFluid/cve-2026-61732-labTrust ok 1000 stars
Benign, self-contained reproduction of CVE-2026-61732 (Decepticon ChatML role-boundary forgery)
created 25 Sep 2026
Fake proof-of-concept repositories have been used to deliver malware to the researchers who run them. The trust verdict is our engine's heuristic over the repository (account age, what it ships, how its stars arrived); it is a reason to read the code first, not a finding about its author.
References
Watch this one?
KEV additions reach Telegram and the KEV-only RSS as they land.