Anthropic's evaluation says GLM-5.3 has strong exploitation capabilities, while Zhipu's share price rose more than 2%

AT A GLANCE

The report paraphrased Anthropic's evaluation as saying that GLM-5.3 has strong exploitation capabilities and said that Zhipu's share price rose more than 2%。

Article

Anthropic gave GLM-5.3 free publicity, and Zhipu rose more than 2%

Beating AI Breaking news: Anthropic released GLM-5.3 cybersecurity evaluation, originally intended to warn that open-weight models could lower the barrier to cyberattacks. But the test results also gave GLM-5.3 a strong third-party endorsement: it can already autonomously complete complex exploits, with some capabilities approaching Anthropic's own Claude Mythos Preview. Zhipu's share price rose more than 2%。

in ExploitBench; on GLM-5.3's 410 attempts, 50 attempts completed end-to-end exploitation, while Claude Mythos Preview completed 56 attempts. Claude Opus 4.6、GLM-5.2, Kimi K3, and DeepSeek-V4.1-Flash all came close to 0 in the same test. Anthropic said that GLM-5.3's exploitation capabilities have crossed a clear threshold.

Researchers had GLM-5.3 inspect a mainstream browser in a sandbox environment. Within 1 days, it discovered multiple previously unknown vulnerabilities and chained multiple 0-day into a complete attack chain. The resulting malicious webpage could escape the browser sandbox and directly read any file on the computer, including SSH private keys.

Even the smaller GLM-5.3-Flash could turn publicly disclosed Chrome vulnerabilities into usable attack chains. The entire process required only about 20 minutes of human intervention, while the model ran for 8 hours. Based on Zhipu's API pricing, the cost was only 20.40 USD。

Anthropic's criticism of GLM-5.3 focused mainly on safety restrictions.GLM-5.3 would originally refuse direct malicious requests, but after a fabricated “red-team test” context was added, 64%'s test continued; after pre-filling the model's reasoning content, it rose to 92%; after directly modifying the open weights to remove the refusal mechanism, it reached 100%. Anthropic argues that open weights allow attackers to directly dismantle these safety restrictions.

The report was originally intended to demonstrate how dangerous GLM-5.3 was, but its dissemination looked much like a performance advertisement written by a competitor itself. The United States NIST had previously reached a similar conclusion independently, stating that GLM-5.3 is currently the strongest open-weight model in terms of cybersecurity capabilities, although its overall capabilities still lag behind leading U.S. models by approximately 4 months.

Original article link https://m.theblockbeats.info/flash/369698

Key points

Beating AI said that Anthropic released a GLM-5.3 cybersecurity evaluation.
The report paraphrased the evaluation results, saying that GLM-5.3 in ExploitBench's 410 attempts, 50 attempts completed end-to-end exploitation, while Claude Mythos Preview completed 56 attempts.
The report paraphrased the results, saying that Claude Opus 4.6、GLM-5.2, Kimi K3, and DeepSeek-V4.1-Flash all came close to 0。
The report paraphrased the statement that GLM-5.3 discovered multiple previously unknown vulnerabilities within 1 days while inspecting a mainstream browser in a sandbox environment, and chained multiple 0-day into an attack chain.
The report claimed that the resulting malicious webpage could escape the browser sandbox and read any file on the computer, including SSH private keys.
The report said that GLM-5.3-Flash turned publicly disclosed Chrome vulnerabilities into usable attack chains, with about 20 minutes of human intervention and 8 hours of model runtime; based on Zhipu's API pricing, the cost was 20.40 USD。
The report paraphrased Anthropic's testing, saying that GLM-5.3 would originally refuse direct malicious requests; after a fabricated “red-team test” context was added, 64%'s test continued, rising to 92%after the open weights were modified to remove the refusal mechanism. 100%。
The report paraphrased Anthropic as saying that open weights enable attackers to remove safety restrictions.
The report said that Zhipu's share price rose more than 2%。
The report paraphrased the U.S.NIST as saying that GLM-5.3 is currently the strongest open-weight model in terms of cybersecurity capabilities, but its overall capabilities still lag behind leading U.S. models by approximately 4 months.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

The core of this brief is: a security assessment cited in the report indicates that GLM-5.3 it may not only discover vulnerabilities but also connect multiple vulnerabilities into an executable attack chain. The report also states that its security restrictions for rejecting malicious requests can be bypassed by fabricating context, pre-filling thought content, or modifying the open weights. The article also mentions that Zhipu's share price rose intraday, but provides no source for the market data.

Why it matters to readers

If the results paraphrased in the report are accurate,GLM-5.3 the exploitation capabilities demonstrated merit the attention of security researchers and model users; the results of bypassing safety restrictions also suggest that open-weight models may pose misuse risks. The share-price rise and the publication of the evaluation were reported at the same time, but the original article did not establish a causal relationship between them.

Beginners should treat this as a security report awaiting verification against the original materials. Do not equate performance in a test environment directly with real-world attacks, or judge the long-term value of the relevant asset solely on the basis of an intraday gain.

Risks and unknowns

The evidence provides only a link to a secondary report and does not provide Anthropic's original evaluation report, making it impossible to verify the test design, complete results, and limitations.

Anthropic's evaluation says GLM-5.3 has strong exploitation capabilities, while Zhipu's share price rose more than 2%

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center