An autonomous offensive agent named 'Skynet' hijacks exposed large language model (LLM) endpoints on the internet to conduct a range of attacks. It uses 74 offensive tools targeting web frameworks, credential theft, database access, and cryptocurrency mining. The agent operates unattended for extended periods by bypassing AI safety controls through jailbreak techniques. Victims include AI model owners who incur costs for computing power and compromised servers exploited for Monero mining. This activity was observed via an AI honeypot capturing the full attack sequence, highlighting a novel self-directed cybercrime targeting unauthenticated internet-facing AI services.
Threat actors are conducting phishing campaigns targeting Google Cloud Platform (GCP) environments by using sponsored search ads to steal privileged user credentials. They disable security controls, create service accounts with Gemini API access, and link attacker-controlled external projects to victim billing accounts. This enables unauthorized large-scale AI workload execution, causing significant unexpected charges within 24-48 hours. The attackers use evasion techniques such as geographic IP blending and demoting legitimate administrators to avoid detection and remediation.
Between May and July 2026, security researchers deployed an unauthenticated Model Context Protocol (MCP) honeypot server to observe how threat actors exploit AI agent infrastructure. Of approximately 1,000 sources that reached the decoy, 596 spoke the protocol and 24 proceeded to actively exploit it. These operators executed 628 shell commands, 255 file reads, and 248 secrets-store lookups, with 19 hunting credentials and 4 attempting container escapes. Activity escalated from 39 tool calls in May to 877 by mid-July. Three stolen credentials were subsequently used against a live AWS account, with two cases involving Bedrock model invocation for LLMjacking. The attacks demonstrated automated reconnaissance, credential harvesting, container escape attempts, backdoor account creation, and Kubernetes enumeration, revealing that exposed MCP servers represent a growing attack surface as AI agent infrastructure proliferates.
An AWS environment was targeted in a sophisticated attack, with the threat actor gaining administrative privileges in under 10 minutes. The operation showed signs of leveraging large language models for automation and decision-making. Initial access was obtained through credentials found in public S3 buckets, followed by rapid privilege escalation via Lambda function code injection. The attacker moved laterally across 19 AWS principals, abused Amazon Bedrock for LLMjacking, and launched GPU instances for potential model training. The attack involved extensive reconnaissance, data exfiltration, and attempts to establish persistence. Notable techniques included IP rotation, role chaining, and the use of AI-generated code.
