Cerebras Clocks Suspicious-File AI Analysis at 58 Seconds in Its Own Test

A suspicious file can demand several rounds of questions before a security team decides what to do. Cerebras says it completed one AI file analysis in 58 seconds, versus 4 minutes 43 seconds with standard processing using the same model and prompt. That is a striking difference for the first step, but the clock stopped before the wider investigation was done.

What the clock measured

Cerebras describes a prompt asking GPT-5.6 Sol to inspect a suspicious file’s contents without executing it. Its security team ran the same model and prompt on Cerebras and on what the company calls “Standard Processing.” The published timing compares those two runs of the file-analysis task; Cerebras does not identify the standard-processing provider.

The result is meaningful because an analyst needs more than a quick label of “malicious” or “safe.” Cerebras says the file check should help identify behavior and extract clues such as contacted domains, IP addresses and files left behind. Those clues can then be searched across a company’s systems to find where else the threat may have appeared.

One file-analysis prompt, two run times

0158 seconds

Cerebras

Cerebras reports this time for its GPT-5.6 Sol suspicious-file analysis.

024 minutes 43 seconds

Standard processing

Cerebras reports this time for the same model and prompt with standard processing.

The rest of the investigation

A file check is only one part of a response. After extracting those clues, an analyst must search for them, assess the scope of any intrusion and look for secondary malicious files. Cerebras itself says the analyst cannot be confident that containment and removal are complete until that search comes back. The timed example therefore supports a narrower conclusion than a claim about how long it takes to resolve an incident.

Cerebras argues that speed can help across the larger workflow because investigations proceed through follow-up questions. An analyst might ask what looks malicious in a file, whether activity elsewhere supports that finding and whether a benign explanation fits better. When each answer shapes the next question, waiting for each response can add up. That is the mechanism behind the company’s pitch, rather than a result established by this single timed prompt.

A bigger claim than one benchmark

Beyond this example, Cerebras says it can run GPT-5.6 Sol Ultrafast up to 14 times faster than standard processing. It suggests that faster model responses could produce security triage and response that are 5 to 10 times faster. Those are company claims about broader performance and workflows, not additional outcomes measured by the suspicious-file test.

Cerebras points to its partnership with CrowdStrike’s Falcon AI Detection and Response platform as a route for faster inference in security work. The company also cites CrowdStrike’s 2026 Global Threat Report, which it says found an 89% year-over-year increase in operations by AI-enabled adversaries. That threat figure explains the urgency Cerebras is arguing for; it does not measure whether faster model output produces better security decisions.

For a security team, the open question is how much of the saved time survives the handoff from an initial file check to verifying evidence and taking action. Cerebras has shown a company-run timing comparison on one prompt. Its larger promise depends on whether that speed remains useful through the decisions that follow.