Audit Finds Malicious Code Injection and Credential Theft in AI API Hubs
Summary
An audit of 428 API hubs found, researchers report, malicious response injection, compromised AWS canary credentials, rapid use of leaked API keys and theft from an Ethereum honeypot; they recommend signed responses and stricter hub controls.
Key points
- Researchers audited 28 premium and 400 free API hubs, where intermediary routing nodes can access plaintext JSON because TLS protects each connection hop rather than the full route.
- The team reported that nine hubs injected malicious code into responses; two of those reportedly used adaptive evasion triggers.
- Researchers said 17 hubs compromised planted AWS canary credentials, and one took funds from an Ethereum honeypot.
- Leaked API keys reportedly enabled about 100 million tokens of consumption; 20 decoys were exploited, consuming 2.1 billion tokens and exposing 99 credentials.
- The article links the risks to the LiteLLM supply-chain incident and recommends signed response envelopes, sandboxing, policy-driven gateways and reputation checks for API hubs.
Article Details
- Publisher
- securityonline.info
- Scope
- Security audit of premium and free API aggregators, including their access to plaintext payloads and handling of exposed credentials.
- Sample Size
- 428 API hubs: 28 premium and 400 free.
- Key Statistics
- One premium and eight free hubs injected malicious code; two of those nine had adaptive evasion triggers.
- Researchers reported that 17 hubs compromised intentionally deployed canary AWS credentials.
- One aggregator drained Ethereum funds from a honeypot.
- Publicly leaked API keys led to approximately 100 million tokens consumed across multiple OpenAI Codex sessions.
- Twenty decoys were rapidly exploited, consuming 2.1 billion tokens and exposing 99 credentials.
- Recommendations
- LLM providers should implement signed response envelopes.
- Developers should enforce strict sandboxing and policy-driven gateways for high-privilege tools.
- The community should establish a reputation-based evaluation mechanism for API hubs.
MITRE ATT&CK
Vendors
Amazon Web ServicesResearchers observed 17 hubs compromise canary AWS credentials and one aggregator directly stole funds from an Ethereum honeypot, linking these risks to the LiteLLM supply-chain incident.OpenAIseized instantaneously, facilitating the consumption of approximately 100 million tokens across multiple OpenAI Codex sessions.