The most secure infrastructure from training to inference.
AI security agents watch every GPU machine your models run on and fix what they find. Training agents keep improving your models against the benchmarks you set. Your prompts stay encrypted from your app to the GPU.
$ sweep gpu-b300-07 --fix[ssh] scan sshd config, keys and accounts[ssh] ok keys only, passwords off, each login waits for an Admin[ssh] found notebook user becomes root without a password[ports] scan every open port, from outside and inside[ports] found Jupyter on a public address, port 8888[packages] scan installed packages against known vulnerabilities[packages] found CVE-2026-111497, Brotli decoder overflow[packages] found 87 packages with known vulnerabilities[codebase] scan services, cron, import paths and credential files[codebase] found root's Python starts in a directory a tenant can write to[firewall] scan rules, published ports and allowed addresses[firewall] found the console's firewall is not on[hunt] scan past the rules: what restarts, what runs at boot[hunt] found a service named after a font cache restarts a miner[firewall] fixed firewall on, SSH checked from outside before and after[firewall] fixed all other inbound dropped, the ports Docker publishes too[packages] fixed security updates installed by name[ssh] fixed sudo rule changed for the notebook user[codebase] fixed root's Python moved out of the shared directory[hunt] fixed service removed, its check now runs on every machine ✓ scanned again: 7 found, 7 fixed, 0 open ✓ watching for what changes next
| SEVERITY | FINDING | STATE |
|---|---|---|
| LOW | Notebook user becomes root without a password | ✓ FIXED |
| LOW | Jupyter listens on a public address, port 8888 | ✓ FIXED |
| CRITICAL | Brotli decoder overflow, exploited in the wild | ✓ FIXED |
| HIGH | 87 packages with known vulnerabilities | ✓ FIXED |
| HIGH | Root's Python starts in a directory a tenant can write to | ✓ FIXED |
| MEDIUM | The console's firewall is not on | ✓ FIXED |
| HIGH | A service named after a font cache restarts a miner | ✓ FIXED |
- SCANNING:
- CONTINUOUS
- COVERAGE:
- EVERY MACHINE
- THEN:
- THE HUNT
Security agents watch every machine and fix what they find.
AI security agents keep scanning every GPU machine your models run on. Past known vulnerabilities, they look for anything that puts your weights, datasets or checkpoints at risk. Every new finding becomes a check on every machine.
- SET UP BY:
- BLACKBOX
- RUN BY:
- BLACKBOX
- SECURED:
- FROM FIRST LOGIN
Dedicated GPUs, from a single node to a full cluster.
Request a VM, a bare metal machine or a cluster. Blackbox sets up the hardware, runs it and keeps it secured from the first login on.
- VM
GPUs on demand for development, evaluation and smaller runs.
- BARE METAL
Whole nodes with no hypervisor, for full-scale training and serving.
- CLUSTER
256, 512 or 1,024 GPUs in one cluster, for the largest training runs.
~ $ ssh blackbox@203.0.113.7blackbox@b300-07:~$ nvidia-smi --query-gpu=name,memory.total --format=csvname, memory.total [MiB]NVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBNVIDIA B300, 294912 MiBblackbox@b300-07:~$ nproc240blackbox@b300-07:~$ sudo whoamirootblackbox@b300-07:~$ - RUN BY:
- AGENTS
- GPUS:
- DEDICATED
- DATA:
- YOURS
Define the target. Agents keep training toward it.
Define the target and the benchmark that measures it. Agents run the training on dedicated GPUs, read every result and keep improving the recipe, round after round.
- REINFORCEMENT LEARNING
Reward the outcomes your work needs.
- POST-TRAINING
Agents keep a round only if it beats the last checkpoint.
- FINE-TUNING
Adapt an open-weight model to your own data.
65.8 77.0
One job took Qwen3.5-9B this far on JevBench Capability. Microsoft Decision-1 scores 70.8.
HOW A TRAINING JOB RUNS- TRANSIT:
- SEALED
- GPU:
- ATTESTED
- RETENTION:
- ZERO
- AT REST:
- AES-256
No one can read your prompts. Not even us.
Your app encrypts each prompt before it leaves. Only an attested GPU enclave decrypts it. The response is encrypted in the enclave, and only your app can decrypt it.
ENCRYPTED INFERENCEEncrypted end to end. Still the fastest.
Agents tune the inference engine. On the same NVIDIA GPUs, it serves more tokens a minute than other providers. On one B200, it is faster than SambaNova and Groq accelerator chips.
Agentic coding, the same 4 GPUs on both sides. Higher is better.
| PROVIDER | TOKENS/MIN | END-TO-END ENCRYPTION |
|---|---|---|
| BLACKBOX AI4× B300 · 128 sessions | YES | |
| BLACKBOX AI4× B300 · 192 sessions | YES | |
| Fireworks AI4× B300 · 64 sessions | NOT OFFERED | |
| Fireworks AI4× B300 · 96 sessions | NOT OFFERED |
October 2026: one load generator replaying 640 production coding-agent sessions against each dedicated deployment, each at its best operating point. BLACKBOX AI runs two 2-GPU replicas on the same 4× B300. End-to-end encryption: each provider's public documentation, checked November 2026.
- SOC 2AICPA
- ISO 270012022
- ISO/IEC 420012023