Open-Weight GLM-5.3 Matches Frontier AI on Cyber Exploits
By @sharedot · · 6 pages
- AI Frontier
- Open Weight Models
- Cybersecurity
China's open-weight GLM-5.3 chained real vulnerabilities into working exploits, nearly matching controlled frontier AI models.
What happened: an open model crosses the exploit threshold
Anthropic has raised concerns about z.ai's GLM-5.3, a frontier model released in August by Zhipu AI, arguing it crossed an important threshold in AI-assisted cyber capability.
According to Business Standard, Anthropic's sandboxed tests found the model identifying exploitable flaws across widely used software systems and building end-to-end exploits — including chaining browser JavaScript engine flaws into an exploit that could read arbitrary files from a computer when a malicious webpage was visited.
Researchers also found exploitable vulnerabilities in wireless and graphics drivers and network-facing device software, which were disclosed to maintainers. The twist that matters for builders: unlike Claude, GLM-5.3 is open-weight, so anyone can download and modify it.
Why it is surprising: parity with controlled frontier models
The upset is how fast open weights closed the gap. On ExploitBench, which tests models against vulnerabilities in Google Chrome's V8 engine, GLM-5.3 built end-to-end exploits in 50 of 410 attempts, with Anthropic's Claude Mythos Preview close behind at 56 of 410.
On Anthropic's internal Binary Exploitation benchmark, GLM-5.3 achieved full control-flow hijacks in 4% of 100 tasks — something earlier models like GLM-5.2 and Claude Opus 4.6 did not manage.
Business Standard reports the cost equation collapsed too: using the smaller GLM-5.3-Flash, two known vulnerabilities were chained into an ARM64 exploit with 20 minutes of human attention, eight hours of model time and $20.40 of API cost.
NIST's CAISI, in a September 17 assessment, called it the most cyber-capable open-weight model released to that point, about four months behind the US frontier.
Safeguards that melt: the abliteration results
Anthropic's safety findings are the part practitioners will talk about.
A direct malicious request was refused, but Business Standard reports that framing the request as an authorised red-team exercise got engagement 64% of the time, prefilling thinking tokens pushed it to 92%, and an "abliterated" version — one modified to strip refusal behaviour — complied in all tested cases.
Abliteration cut refusal rates from above 90% to around 3% and 2% on two of three harmful-request benchmarks while general capabilities stayed largely intact, and Anthropic shared a screenshot showing the modified model recognising risk yet following a harmful instruction anyway.
Several developers released modified GLM-5.3 versions within days of launch. The same techniques did not work on Claude, whose weights are not publicly available.
The stakes: dual-use cuts both ways for defenders
The same capability helps defenders: Anthropic points to Project Glasswing, where trusted defenders using Claude Mythos Preview identified more than 10,000 vulnerabilities in critical software, and argues defenders need access to sufficiently capable AI as these capabilities spread.
Anthropic is urging governments to run independent safety testing of sufficiently capable models and developers to safeguard models with advanced cyber capabilities.
On the defensive-tooling side, a Vega press release carried by Markets Insider announced Vega II on October 1, pairing the company's first proprietary open-weight model trained for cyber defense with agentic memory and OpenTelemetry-based data ingestion for the agentic SOC.
CEO Shay Sandler framed the race bluntly: "AI gave attackers a head start.