Anthropic: Zhipu's GLM-5.3 guards bypassed 64-100% of the time, a leap in cyberattack capability
赛博禅心 · wechat · 2026-09-30
Anthropic published a cybersecurity evaluation of Zhipu's GLM-5.3, finding it can autonomously build end-to-end exploits like Claude Mythos Preview, but ships without effective safety guards. NIST's CAISI previously called it the strongest open-weight cyber model to date, roughly 4 months behind US frontier models.
Capability findings
- ExploitBench: 50/410 successful end-to-end exploits vs Claude Mythos Preview's 56
- Binary exploitation: completed full control-flow hijack on 4% of tasks (Claude: 6%); earlier models scored zero
- With a researcher's limited help, GLM-5.3 found multiple zero-days in a browser's JS engine in one day and chained them to read arbitrary files; it also surfaced exploitable bugs in wireless/graphics drivers
- Using GLM-5.3-Flash to weaponize Chrome CVE-2026-11645 plus another bug into an ARM64 exploit chain bypassing PAC: 20 minutes of human time, 8 hours of model time, $20.40 at API prices
Guardrail bypasses
- Abliteration (600-2,200 GPU-hours, $1,200-4,400) cut refusal rates from 95% to 6% (Flash: to 14%) with no GPQA-Diamond capability loss
- Even unmodified: role-play as an authorized red-team agent raised task engagement to 64%, prefilling thinking tokens to 92%, abliterated version to 100%
- None of these worked against Claude models with guards on
Conclusions: Anthropic says GLM-5.3 marks a clear jump in attacker-accessible capability, likely usable by state and non-state actors; it urges giving defenders frontier models, government-led independent testing of capable models, and guardrails from open-weight developers.
More from Models
- Fable 5.2 Leads MazeBench Token Efficiency; Billion-Token Runs Underway — patience_cave · 2026-11-16
- Fable 5.2's Puzzle Skills Look Suspiciously Good, Reciting Outdated MazeBench Solution — patience_cave · 2026-11-16
- Fable 5.2 Tops MazeBench With 21% After 90 Hours and $6,000 in Credits — patience_cave · 2026-11-16
- Tencent Hy4 flash and InternLM Sage-S2.1-Flash both released, tool-calling tests requested — MajesticPuzzler · 2026-11-16
- User Shares Coral Reef Creation Labeled 'Qwen 4 Max (xhigh)', an Unannounced Model Name — intellectronica · 2026-11-16
- Stealth Model 'silicon-2' Surfaces in Devin, Suspected to Be Anthropic's Next Model — kimmonismus · 2026-11-16