Summary
- Chinese AI firm Zhipu says its GLM-5.3 model outperforms Anthropic’s and OpenAI’s equivalents on the CyberGym cybersecurity benchmark for vulnerability discovery.
- GLM-5.3 reportedly found 2,436 vulnerabilities across 269 real-world codebases, including 1,097 medium-to-high severity issues, some dormant for decades.
- TCS and HCL both disclosed employee data exposures via stock exchange filings; both say customer data is unaffected and the stolen data is several years old.
- TCS’s attacker reportedly used password spraying and MFA fatigue — techniques the company says it has defended against for more than two years, implying the breach predates those controls.
- HCL’s investigation remains ongoing.
A capable bug-finder emerges from China
Zhipu, a Chinese AI company, last week released GLM-5.3 and accompanied the launch with benchmark data claiming the model surpasses comparable offerings from Anthropic and OpenAI on CyberGym — a benchmark designed to test a model’s ability to solve real-world cybersecurity challenges. According to Zhipu, GLM-5.3 beat models identified as Fable 5 and GPT-5.6 Sol on that specific test. The company was candid that GLM-5.3 performs less well than Western models on other security and coding benchmarks, which is worth noting before drawing sweeping conclusions.
Reasoning across exploitation chains, not just spotting flaws
What distinguishes Zhipu’s claims is not simply that the model identifies vulnerabilities in isolation. The company wrote that GLM-5.3 “began to reason across multiple stages of exploitation, forming coherent plans for complete exploitation chains.” That is a meaningful distinction for defenders: a model capable of chaining vulnerabilities into a complete attack path is a more serious offensive tool than one that flags discrete weaknesses. Zhipu says it tested the model against real codebases in partnership with Chinese companies, surfacing 2,436 vulnerabilities across 269 projects — 1,097 of them rated medium to high severity — spanning kernels, browser engines, web applications, network protocols and open-source infrastructure. The company notes some of these issues had gone undetected for decades, with the oldest tracing back roughly 40 years.
Strategic implications for the security community
The speed of development is the signal CISOs should register. Zhipu notes that cyber capability scaled faster than it anticipated during post-training. The Register’s analysis frames it plainly: any advantage the United States held by virtue of being home to Anthropic has largely dissipated. Whether or not GLM-5.3’s benchmark results hold up to independent scrutiny — and no corroborating source has yet validated them — the trajectory is clear. AI-assisted vulnerability discovery is no longer a capability exclusive to Western-aligned research environments. Security teams should factor that into their threat modelling, particularly for long-lived open-source components and legacy codebases.
TCS and HCL disclose employee data exposures
Separately, two of India’s largest IT services firms — Tata Consultancy Services and HCL Technologies — both filed stock exchange disclosures last week acknowledging that employee data had been exposed. TCS said it received threat-intelligence alerts about potential exposure of employee information and investigated, finding no credible evidence of a breach of its own systems or customer environments. The data involved is described as basic employee information more than four years old. The attacker reportedly used password spraying and MFA fatigue attacks. TCS noted it has had strong safeguards against both techniques for more than two years — a statement that implicitly places the incident before those defences were introduced.
HCL’s investigation continues
HCL’s disclosure, also via a stock exchange filing, acknowledged claims from a hacker group regarding exposure of a limited set of employee data described as several years old. HCL similarly stated that customer data is safe. Unlike TCS, HCL indicated its investigation is still underway, meaning the full scope remains undetermined. For security leaders who rely on either firm as a managed service provider or systems integrator, the assurances regarding customer data are notable, but it is reasonable to seek direct confirmation from your account teams given the investigations are at different stages of completion.
Why it matters
GLM-5.3’s reported capability to reason across multi-stage exploitation chains — not merely flag individual bugs — represents a qualitative shift in what AI-assisted offensive tooling may look like in practice. If the benchmark results reflect real-world performance, defenders face a situation where capable vulnerability discovery tools are no longer the exclusive domain of well-resourced Western actors. Simultaneously, the TCS and HCL disclosures serve as a reminder that MFA fatigue and password spraying remain effective against organisations that have not consistently enforced modern authentication controls across all systems. Both stories point to the same underlying discipline: reduce the attack surface in your legacy code and harden authentication before an automated system — or a determined attacker — finds what you have not.
What to do now
- Review your organisation’s exposure to long-lived open-source dependencies and legacy codebases, particularly those that may not have received regular security audits.
- If you use TCS or HCL as a service provider, request a direct briefing from your account team on the scope of their respective incidents and confirm your data environments are unaffected.
- Audit your MFA implementation to ensure it is resistant to fatigue attacks — push-based MFA without number matching or phishing-resistant alternatives (such as FIDO2) remains a meaningful risk.
- Assess whether AI-assisted code scanning tools are part of your vulnerability management programme, given the growing accessibility of such capabilities to a wider range of threat actors.
- Revisit threat models that assume AI-assisted exploitation is limited to well-resourced nation-state actors, as the barrier to entry appears to be lowering.
