Lambert Sticks to Open-Weight Cyber Call Despite Black October

Nathan Lambert's October check-in on his own cyber risk call: he still thinks the open-weight disaster framing was overstated, but less confidently than he did in April. In a post titled "Was I wrong about open models & cyber risks?" on Interconnects, Lambert returns to his earlier argument that if Anthropic's Claude Mythos had leaked, "the world would have been more or less fine," and holds the line while flagging what "Black October" changed about his certainty.

The context is a run of critical disclosures Lambert groups under that name, including PHP2Shell and GitLab. He treats the vulnerabilities as real and systemic without accepting the step-change framing April's hawks used. He points to an Epoch AI demonstration in which, as the researchers wrote, "We used GLM 5.3 to create a website that, simply when visited by a Firefox user and without further interaction, gave the attacker the ability to exfiltrate files from the user's computer." His read is that current open weights are already weaponizable enough to study, which is a different claim from a frontier release breaking things categorically.

Lambert names the April voices he's measuring himself against. Rob Wiblin on 80,000 Hours said that if Mythos "fell into the hands of criminals or hostile state cyber actors, it would be an actual disaster." Anthropic's Project Glasswing announcement warned capabilities could proliferate "beyond actors who are committed to deploying them safely." Zvi Mowshowitz wrote that an open Mythos "would give attackers a cornucopia of zero-day exploits for essentially all the software on Earth." Lambert's counter is partly operational, with closed models heavily involved in recent hacks, and partly economic, with open-weight adoption still gated by serving costs and compute scarcity.

He borrows from a malware researcher he quotes: "Life is a lot easier once you stop seeing threat actors as hackers who want to break into systems by any means possible, and start seeing them as businesses with well defined business processes." Attackers follow margin, not model releases.

What Lambert does concede is cost. He expects "a forced industry update to a new way of working with software, which will clearly be expensive and accompanied by some hard work," and frames the longer arc as "working through the backlog of cyber vulnerabilities as we transition from the human-focused internet to closed silos of human activity mediated by agents." His one-line verdict on himself: "I largely still stand by my above writing a few weeks ago, but I'm nervous about it."