Open-weight AI models close the gap on multi-step network attacks
The cost gap is the sharper signal: an attack that runs $17 to $46 on a closed model costs $2 to $3 on an open one.
Vals, a company that measures AI capabilities, tested how well models could complete multi-step attacks on simulated corporate networks, and found open-weight models now close to the best closed ones. OpenAI's Daybreak, limited to trusted partners, completed 81 percent of tasks on average, while DeepSeek, an open-weight model, completed 73 percent. Running the attack was far cheaper on the open model, at $2 to $3 in tokens against $17 to $46 for GPT-6-Sol. For defenders this weakens the case that gated access will keep advanced hacking capability out of reach for long. The reporting also notes that real-world cybercriminal and state-backed activity has mostly relied on older, less capable models, a pattern the article's own observation supports rather than any expert quoted in it.
Why it matters
For people watching AI security, the finding weakens the idea that restricting the most capable models to trusted partners will keep advanced hacking out of reach for long. Open-weight models now complete simulated multi-step attacks almost as well as the best closed ones, and at a fraction of the price. Real-world attackers have so far relied mostly on older, less capable models, so the threat has not yet materialised at this level, but the gap in capability and cost is what changes the picture.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.