Can AI Growth Absorb High Rates? A 18:25 한국어
Jump to the 3 latest entries Google Project Suncatcher — first Gemma 4 response from an orbital TPU, a 1kW proof of concept(1)Anthropic's autonomous red team finds monitor-evading misalignment datasets in every run(2)Epoch AI — open-weight GLM 5.3 exploits the patch gap in all six Firefox releases (about $1,000 per version)(2)

← Back to the record

Epoch AI — open-weight GLM 5.3 exploits the patch gap in all six Firefox releases (about $1,000 per version)

Date2026-10-29
CategoryFrontier Models
JudgementRefutes
Gauge weight2 / 3
SourceEpoch AI

Summary

In an evaluation Epoch AI published on October 29, 2026, Z.ai's open-weight model GLM 5.3 used the patch gap to compromise all six Firefox releases shipped between September 17 and October 26 (156.0, 156.0.1, 157.0, 157.0.1, 158.0, 158.0.1); once the harness was ready, the marginal cost was about $1,000 per version, the time budget was up to 48 hours per version, and the fastest success took 12 hours. The model also rediscovered an undisclosed sandbox-escape vulnerability on every attempt and chained it; that vulnerability was reported to Mozilla and patched in Firefox 159.0 (October 27). No success rate was disclosed.

A/B Impact Analysis

Filed on the refuting side of the Scenario A thesis: an open-weight model with no guardrails compromising a major browser for about $1,000 per version strengthens the security and regulatory friction that acts as delay pressure. The run was recorded as local self-correction under an external oracle, but judged not to meet a standard requiring single-agent survival without auto-resume scaffolding: no success rate was disclosed, the structure ran until success, and the harness was built by humans.

← Previous entryNext entry →