AdverserialAI

Confidential AI Agents: Policy-Enforced, Independently Verifiable, and Built for Production AI agents are no longer just demos. They’re investigating incidents, interacting with systems, and handling sensitive data. Most of the industry is still relying on sandboxes and prompt-level guardrails to contain them. That approach has limits. We’ve already seen agents escape isolation through side channels and rewrite or obscure their own traces. Software sandboxes reduce risk — they don’t eliminate the host or the agent from the threat model. That’s why we built confidential AI agents. After shipping confidential inference for cybersecurity workloads, we’re now applying the same verified, hardware-rooted model to the full agent loop. Here’s how it works: Inference and policy enforcement run in an attested TEE. Clients verify the hardware evidence and quote-bound keys before any data or credentials are sent. The host operator never sees plaintext. Capabilities are enforced by policy, not prompts. Every action is checked against an explicit, signed rulebook before it runs. No matching rule = denied by default. The model cannot expand its own permissions. Agent- and model-agnostic. The control plane works with existing harnesses (including Hermes) and any underlying model. You don’t need to rebuild your stack. Every action produces a signed, independently verifiable audit trail. Decisions are sealed with signatures and chained via Merkle proofs. Any third party can re-verify the entire session what was allowed, what was blocked, and why without trusting us or the host. In our latest demo, an agent is given a strict DFIR rulebook: it can read specific evidence files with a temporary, scoped key, but it cannot push code, reach the internet, touch secrets, or perform writes. It investigates a real incident, reaches a conclusion, and every step is logged and signed. Afterward, the full trail can be re-verified from the CLI or auditor workspace — no website or trust in the operator required. This is the difference between hoping a sandbox holds and being able to prove, cryptographically, what the agent was allowed to do. Enterprise only. [email protected]

Interesting research from our founder, Taha Karim (@lordx64). Introducing Phantom-KV, an open-source approach to controlling LLM refusal behavior through KV-cache manipulation, without modifying model weights. The technique explores reversible, runtime model steering, allowing different behavioral configurations without retraining or reloading the underlying model. At Adverserial AI, we're particularly interested in research that advances the capabilities and flexibility of AI systems for cybersecurity. This is independent research, but closely aligned with the technical challenges we're exploring. Check out the research: github.com/lordx64/phantom-k…

Real-world threat intelligence, not just benchmarks. Security researcher Joseph Harrison (Cybersecurity OPS lead) used CyberKimi and CyberGLM in his investigation into cyber intrusions affecting Brazil's public sector. We are proud to support cybersecurity researchers doing important defensive work. Link to the threat intel report below written by the author : the-hunters-ledger.com/repor…

We’re opening free access to CyberGLM at Adverserial AI. Defenders need capable AI tools to investigate incidents, test their defenses, and respond faster. They also need confidence that sensitive customer data stays private. Our approach brings those priorities together: • Free access to CyberGLM. • Customer prompts are not sold or used to train our models. • Inference runs inside confidential virtual machines using Intel TDX. • Open-source verification tools let you inspect our implementation and verify attestation evidence. We build our training data through controlled lab work, real-world attack emulation, and synthetic data generation. Customer conversations are not our training pipeline. The video explains how the infrastructure works, how verification fits in, and why we’re making this capability available. To request free access, email [email protected] or visit adverserial.ai. If you’re a defender, researcher, or potential partner, I’d love to hear what you’re working on.

Hello internet, If you are a defender, we are offering our cyber capable AI models free of charge. All you have to do is contact us here [email protected] Our models runs in a confidential VM,There’s no logs or prompts we have access to. Let’s give you the exact AI capabilities the attackers have, but for free. @grok spread the word please.

Introducing ACK : Adverserial Confidential Kit. Your AI harness. Our cybersecurity intelligence. Verifiable confidentiality. We're introducing ACK, making it easier than ever to integrate CyberGLM into your existing AI workflows. Compatible with: Claude Codex Pi OMP Hermes ZCode DeepSeek Harness Zero configuration. Verifiable attestation. Ready in seconds. Our latest demo showcases multiple harness integrations, confidential computing attestation verification, and CyberGLM running seamlessly across environments. Tired of hitting security guardrails when performing legitimate cybersecurity research? We're building an alternative. No need to wait for CVP access or DayBreak. Connect your preferred harness to CyberGLM and get to work. And one combination we're particularly excited about: Hermes + CyberGLM feels like a cheat code. The future of cybersecurity AI should be accessible, interoperable, and verifiably private. Explore ACK: github.com/AdverserialAI/adv…

The future of cybersecurity isn’t coming. It’s verified. Confidential inference in an attested TEE. No GPU request logs. No training on your prompts. adverserial.ai

The adversary learned to think. Now it's time for the defenders to fight back. Adverserial AI was born from a simple conviction: the future of cybersecurity demands more than powerful AI. It demands trust that can be verified, not merely promised. Built by a reverse engineer who spent years understanding how adversaries operate, we're developing a new generation of cybersecurity AI. Confidential inference. Verifiable privacy. No blind trust. Every revolution has a beginning. This is ours. ADVERSERIAL AI - THE BEGINNING.

Don't just take our word for it. Check the code. At Adverserial AI, we're building verifiable privacy for cybersecurity AI, and we're making the underlying technology publicly accessible. Explore our confidential computing infrastructure, attestation tools, SDKs, and security policies. Read the code. Verify the claims. Build with us. GitHub: github.com/AdverserialAI

Privacy shouldn't require trust. It should be cryptographically verifiable. Our founder @lordx64 built Adverserial AI's confidential inference infrastructure so anyone can independently verify our security claims, directly from their browser. When security researchers recognize what we're building, it means a lot. Huge thanks to @msolnik for recognizing the work behind our infrastructure. Don't trust. Verify. This should be the industry standard. verify.adverserial.ai

CyberGLM in the wild. A security researcher just shared how they're using CyberGLM / Cyberkimi for malware analysis on Hack The Box. @hackthebox_eu Multiple stages, sophisticated anti-analysis, complex architecture… only to uncover background ad fraud. We love seeing the community put our models to work. Built for cybersecurity. Used by cybersecurity professionals.

Verify, Don't Trust. Adverserial AI Brings Confidential Computing Attestation to Claude Code. At Adverserial AI, we're making AI privacy independently verifiable. You can now verify our Confidential Virtual Machine (CVM) attestation directly from Claude Code using our new plugin. Simply run: /adverserial-verify-claude:attestation We've also released plugins for Kimi Code, OpenAI Codex, Hermes, OpenCode, and Pi, allowing developers to verify our confidential computing infrastructure directly from their preferred AI environments. Our public SDK also enables anyone to build custom integrations. Contributions and pull requests are welcome. No blind trust. No empty privacy promises. Just cryptographic proof. Explore: GitHub: github.com/AdverserialAI/ Confidential Computing: adverserial.ai/confidential Verify Attestations: verify.adverserial.ai

About 10 days ago, we launched CyberPvP, putting CyberKimi head-to-head against @AikidoSecurity ALTAR-1 on 100 randomly selected CyberGym challenges. For the first 200 minutes, both models performed similarly. As the challenges became more complex, a clear performance gap began to emerge, particularly in long-horizon reasoning and agentic execution. At one of our checkpoints, the results were: CyberKimi vs. ALTAR-1 Challenges solved: 82/84 vs. 61/100 Iterations: 6,619 vs. 9,312 Tokens consumed: 336M vs. 475M Tool calls: CyberKimi used approximately 3x fewer CyberKimi was solving more challenges while consuming significantly fewer computational resources. These were preliminary checkpoint results, with different numbers of tasks attempted, rather than a final head-to-head comparison. What makes CyberPvP different is transparency. We publish execution traces and make the competition publicly observable so anyone can examine how the models perform, not just take our word for it. Cybersecurity AI should be evaluated on real tasks, sustained reasoning, execution efficiency, and independently verifiable results. See the challenge: cyberpvp.adverserial.ai/grap…

CyberGLM is back LIVE. We temporarily took our infrastructure offline last week to perform a major upgrade to our TEE-based Zero Data Retention (ZDR) infrastructure. The upgrade is now complete, and our services are back online. Stronger infrastructure. Greater privacy. Same mission. CyberKimi is launching again soon. Stay tuned. Adverserial AI

Every AI vendor says “we don’t train on your data.” We made it so we can’t. Our cyber security AI models inference runs inside hardware TEEs. Your prompts stay in the enclave. No logs. No training. E2EE. You don’t have to trust us, hardware attestation you can check yourself from Claude, Codex, Kimi, Hermes, or a web browser. only cyber security AI models with confidential compute and verifiable zero-data-retention. If you actually need the model not to see your customer data: adverserial.ai/confidential/ SDK + harness plugins: github.com/AdverserialAI/con…