What Landed
Simon Willison flagged Drew Breunig's "nesformer": a working NES 6502 CPU emulator trained into the weights of Qwen3.5-4B. It runs real ROMs very slowly. Opus 5.5 built it from two short prompts, using a 15-year-old public 6502 conformance suite to check correctness. Separately, Nathan Lambert revisited his claim that an open-weight Mythos would have been "fine." GLM-5.3, an open-weight model at Mythos-class capability, has produced no visible step change in harm in its first month, though "Black October" and Epoch AI's GLM-5.3 Firefox exploit show the cyber risk is real and accelerating.
Why It Matters
Nesformer is a toy, but how it was built matters. Agents get from a vague prompt to a verified result when an executable spec already exists. For enterprise migrations, ports and reimplementations, test coverage decides what an agent can do more than model choice does. Lambert's argument moves the cyber question away from open versus closed weights and toward operations. Disclosures are rising exponentially and patching costs are climbing, and the open-model abuse path he considers most plausible is a poorly secured model API. For an advisor on private AI, both points lead to practical deliverables, not policy positions.
Worth Raising With Customers
- Conformance suites before agent-driven migrations. Before scoping agent work on legacy ports, parsers or format converters, check whether an executable test oracle exists. If it doesn't, building one is the first deliverable.
- Harden self-hosted inference endpoints. Exposed open-model APIs are the likeliest open-model attack vector Lambert names. Audit authentication, network exposure and rate limits on any internal vLLM, Ollama or similar deployment.
- Budget for higher patch volume. Treat the next period as sustained growth in vulnerability disclosures rather than a one-off spike. Check that patching and response capacity can absorb it, and treat model weights from outside sources as opaque, untrusted artifacts.