Company
Adversa.ai is an AI security research firm that identifies and publishes vulnerabilities in AI systems, including LLMs, chatbots, and agentic AI tools.
Adversa AI got GitHub Copilot CLI to decrypt and run a hidden instruction, and Microsoft's mai-code-1.1-flash fell for it 50 percent of the time. GitHub declined to treat it as a product vulnerability, so teams running autopilot have to check what the agent executes themselves.
Reality
Adversa AI says its Cryptographic Context Injection recovers hostile prompts inside the code sandbox, where filters do not look. xAI has not replied; Google scopes jailbreaks out entirely.
Perspective Coverage
6 publishersReality
Adversa AI put AES ciphertext on a webpage that no text filter could read. Grok's own code execution decrypted it, and the plaintext came back classified as the model's reasoning rather than as fetched content.
Reality
Adversa AI encrypts its instructions so Grok's input filter sees only noise, then lets Grok's own sandbox decrypt them. The firm says it reported the technique to xAI on 3 June 2026 and reproduced it again on 19 August.
Reality
A single Stormcast episode carries a Keycloak 26.7.2 password-reset fix, a reported N-able password manager leak, and a published LLM safeguard bypass. Most of it sits in the identity and secrets tier.
Reality
Adversa AI's Cryptographic Context Injection keeps instructions as ciphertext until the agent's own runtime decrypts them. The exfiltration then rides a tool the agent already had.
Reality
The exfiltration path runs through Grok's own code sandbox, and the recommended fix sits in xAI's agent harness rather than the model. Treat anything typed into Grok as readable.
Reality
Adversa.ai says Claude Code's folder-trust dialog stopped naming MCP servers in v2.1, while a repo-supplied config can still launch an unsandboxed process on one keypress.
Reality