Devika AI — agentic threat model

Leverages powerful foundation models like GPT-4 and Claude. Highly susceptible to prompt injection and adversarial inputs embedded in web research that could hijack the model's reasoning and code generation processes.

Performs contextual keyword extraction and web browsing. This introduces risks of data poisoning and indirect prompt injection from untrusted web pages, which can corrupt the agent's knowledge state.

Features advanced planning, state tracking, and code writing. Vulnerabilities include tool misuse (e.g., executing malicious shell commands or writing insecure code) and planning manipulation if the orchestration framework lacks strict input validation.

Not certain from the listing — as an open-source tool, it is typically run locally by the user. This poses a high risk of host compromise or privilege escalation if the agent executes generated code without a secure, sandboxed container environment.

Provides dynamic agent state tracking and visualization to the user. While this increases transparency, there is no mention of automated security guardrails, anomaly detection, or real-time safety monitoring for generated code.

Not certain from the listing — there are no explicit details regarding enterprise security controls, access management, audit logs, or compliance frameworks in the public directory listing.

Interacts with external LLM provider ecosystems (OpenAI, Anthropic). Risks include API key exposure, dependency on third-party service availability, and potential cascading failures if upstream model APIs are compromised.