CVE-2026-61732 in Decepticon
Summary
by MITRE • 09/24/2026
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.
If you want to get the best quality for vulnerability data then you always have to consider VulDB.
Analysis
by VulDB Data Team • 09/24/2026
The vulnerability identified within versions of the Decepticon autonomous hacking agent prior to release 1.1.17 represents a critical failure in input sanitization that leads to prompt injection and subsequent arbitrary code execution. This flaw is rooted in how the application processes data retrieved during its reconnaissance phase, specifically when wrapping web crawl results into messages for interaction with Large Language Models. The core technical deficiency lies in the lack of neutralization for ChatML special-token literals within user-supplied content. In many modern LLM architectures, particularly those utilizing chat-based interfaces like OpenAI's API or open-source equivalents such as vLLM and Ollama, specific token sequences are reserved to denote structural boundaries between different roles, such as system instructions, assistant responses, and user inputs. When these special tokens are present in raw data without proper escaping or filtering, they can be misinterpreted by the model parser as legitimate structural markers rather than literal text content.
Under the Bring Your Own Key deployment model supported by Decepticon, users configure credentials for any OpenAI-compatible endpoint. A significant portion of open-source and self-deployed LLM providers do not filter special-token literals from user-provided content in their default configurations. This architectural assumption creates a dangerous attack surface where an adversary can plant specific strings within target web pages during the crawling process. When Decepticon ingests this data, it passes these unneutralized tokens directly to the language model. The LLM then parses these injected sequences as authoritative role-boundary token IDs, effectively forging a new operator turn that appears legitimate to the system. This manipulation allows an attacker to bypass the agent's internal guardrails and security constraints, which are designed to limit the scope of autonomous actions within its operational environment.
The operational impact of this vulnerability is severe, resulting in arbitrary command execution inside the Kali Linux sandbox where Decepticon operates. By successfully injecting these structural tokens, an attacker can manipulate the LLM into executing commands that were not intended by the red team operator or the system's safety protocols. This constitutes a classic case of prompt injection leading to remote code execution, undermining the integrity and security boundaries established for autonomous agents. The vulnerability aligns with CWE-78 Improper Neutralization of Special Elements used in an OS Command (OS Command Injection) as it allows external input to influence command structure, and also relates to CWE-1386 Insecure Prompt Injection where untrusted data is processed without adequate validation before being fed into a generative model. From the perspective of the MITRE ATT&CK framework, this behavior corresponds to techniques involving prompt injection for privilege escalation or defense evasion, allowing an adversary to subvert automated security controls and gain unauthorized access to underlying system resources.
Mitigation strategies must focus on rigorous input sanitization at both the application level and potentially through configuration adjustments in the LLM provider if feasible. For Decepticon users, upgrading to version 1.1.17 or later is essential as this release patches the issue by properly neutralizing ChatML special-token literals before they are passed to the language model. This ensures that any such characters found within web crawl results are treated strictly as data rather than structural commands. Additionally, organizations employing autonomous agents should implement strict allow-listing for permitted inputs and consider deploying LLM providers with enhanced security configurations that filter or escape potentially dangerous tokens by default. Regular auditing of third-party integrations and continuous monitoring for anomalous model behavior can further reduce the risk associated with prompt injection vulnerabilities in automated red team tools.