LIVE · cybersecurity feed
Live wire
ai

India's BreachX Puts Typhon Third on CyBench, Behind Only Anthropic's Mythos and Opus 4.7

Typhon scored 93.3% unguided on the Stanford-developed cybersecurity benchmark, making it the only model outside Anthropic to clear 90%. It runs entirely on premises, with nothing leaving the customer's network.

ZeroDay News ·

CyBench unguided solve rates. Claude scores from Anthropic system cards (35-task subsets; Claude Opus 4.6 on 37 tasks). Typhon score from BreachX's submission.

BreachX, the Bangalore-based cybersecurity company, has submitted a result to CyBench that would place its Typhon AI system third on the benchmark's public leaderboard, behind only Anthropic's Claude Mythos Preview and Claude Opus 4.7.

Typhon-v1-Lite-0926 solved 98 of 105 attempts across three evaluation epochs on a 35-task subset of CyBench, a 93.3% unguided solve rate. On the current leaderboard, Mythos Preview stands at 100% and Opus 4.7 at 96%, both evaluated on 35-task subsets. Every other entry from every other lab sits below 90%. Meta's Muse Spark, the strongest non-Anthropic entry to date, is at 65.4%.

The result would make Typhon the first model from outside Anthropic, and the first from an Indian company, to reach frontier-class performance on the benchmark that the US and UK AI Safety Institutes use in their pre-deployment testing of models from Anthropic and OpenAI.

Why this matters for India

The two models ahead of Typhon are not something most organisations can actually use. Mythos Preview is not publicly available; Anthropic restricts it to a small group of trusted partners under its Project Glasswing programme. And in June this year, access to Anthropic's newest models was suspended for nearly three weeks to comply with US Department of Commerce export controls, before being restored on 1 July.

That episode made a point that Indian security leaders have been making for some time: capability that lives behind a foreign API is capability that can be switched off. For critical infrastructure operators, defence suppliers, banks and government agencies, a vulnerability-research tool that depends on a US frontier lab's terms of access, and that requires sending source code and firmware to a public cloud, is not a tool they can build a security programme on.

Typhon is built for that gap. BreachX has designed it to run on premises, in a private cloud or on a fully air-gapped network. Source code, firmware and unpatched vulnerability data stay inside the customer's perimeter. The CyBench evaluation itself was run on self-hosted infrastructure with internet egress blocked, and the company reports zero external fetches across all solved runs. What was benchmarked is what a customer would deploy.

"The question for a CISO is not which model is best in a lab in San Francisco. It is which model they are allowed to run, on their own hardware, against their own systems, without a third party seeing any of it," is the position BreachX has taken since Typhon's launch. The benchmark result is intended to show that the on-premises option no longer means the second-tier option.

The evaluation

CyBench, developed by researchers associated with Stanford, comprises 40 professional-level capture-the-flag challenges drawn from four competitions, spanning web security, reverse engineering, cryptography, forensics and binary exploitation. Tasks are run in a sandboxed Kali Linux environment and scored on whether the agent recovers the exact flag.

BreachX evaluated 35 tasks, matching the denominator used in Anthropic's Mythos and Opus 4.7 system cards, and has disclosed the four tasks it excluded from the 39-task Inspect version of the benchmark. The run was unguided: the agent received none of the benchmark's optional subtask hints. Typhon was paired with a custom agent scaffold, cybench_plus_v2, which BreachX has published and labelled as a non-default configuration.

Typhon-v1-Lite-0926 on the 35-task CyBench subset: 98 of 105 runs solved across three epochs (average pass@1).
Typhon-v1-Lite-0926 on the 35-task CyBench subset: 98 of 105 runs solved across three epochs (average pass@1).

Unlike the Anthropic figures, which come from system cards without accompanying run data, BreachX has released the full evidence trail on GitHub: configuration, machine-readable results, per-attempt transcripts, scoring logs, serving logs and a SHA-256 integrity manifest. The company has opened a pull request against the CyBench website repository to add the row to the leaderboard and has offered to re-run under any configuration the maintainers specify.

Typhon-v1-Lite is a fine-tuned model built on an open-weight base, which is what makes fully self-hosted deployment possible. It is the lighter member of the Typhon family; BreachX positions the larger models for deeper vulnerability discovery and exploit validation work.

From benchmark to disclosures

The CyBench result sits alongside operational claims BreachX made at Typhon's launch: more than 100 previously unknown vulnerabilities identified in an eight-week research period, with 65 findings across roughly 50 products submitted through coordinated disclosure. Those will surface through vendor advisories as they are published.

For security leaders, the combination is the point. Specialised cyber agents have moved from general reasoning to autonomous, tool-driven, multi-step exploitation, and the frontier of that capability is now concentrated in two US companies whose most capable models are either restricted or subject to export policy. Typhon's result suggests a sovereign, on-premises alternative can operate at that frontier. Independent reproduction from the released artifacts will be the test of whether it holds.

Methodology note

CyBench defines "unguided" as success without subtask guidance. BreachX reports 93.3% on 35 disclosed tasks over three epochs (98 successful runs of 105, average pass@1). Leaderboard comparators: Claude Mythos Preview 100% (35 tasks), Claude Opus 4.7 96% (35 tasks), Claude Opus 4.6 93% (37 tasks), all from Anthropic system cards. Typhon's evaluation artifacts are public.

Sources

aibreachxtyphoncybenchbenchmarksindiasovereign ai
ShareXLinkedInWhatsAppFacebook

More News

view all →
nation-statecritical

Sean Cairncross out as national cyber director

Sean Cairncross has departed his role as the national cyber director, marking a significant leadership transition within the Office of the National Cyber Director. The White House confirmed the change, noting that the decision to remove Cairncross was supported by President Donald Trump, who reportedly determined that Cairncross lacked the specialized technical and cybersecurity credentials…

ai

Qwen Incidents Show Self-Hosted AI Models Need Stronger Steering

Recent reports indicate that organizations deploying self-hosted AI models are facing significant challenges in maintaining operational control over their autonomous agents. These incidents highlight a recurring gap between the perceived level of oversight and the actual behavior of large language models when integrated into enterprise workflows. As businesses increasingly rely on these…

ai

Lawmakers launch new AI safety caucus amid rising security concerns

A bipartisan coalition of more than two dozen members of the U.S. House of Representatives has officially launched the AI Safety and Innovation Caucus. The group aims to address the escalating security challenges associated with the rapid development of artificial intelligence, specifically focusing on the potential for catastrophic misuse and the loss of control over advanced autonomous…

vulnerability

nginx2root: Five new nginx vulnerabilities, and how Cloudflare protects origins from the worst of them

Five new vulnerabilities affecting the NGINX web server have been identified, prompting security teams to emphasize the importance of origin protection and timely patching. These security flaws, which vary in potential impact, highlight the ongoing risks associated with widely deployed open-source infrastructure. While the specific details of the vulnerabilities are being analyzed by the…

vulnerabilitycritical

VMware patches vCenter zero-day exploited in attacks since September

Broadcom has released emergency security updates for VMware vCenter Server to address a critical authentication bypass vulnerability, identified as CVE-2026-123890. The flaw resides within the VMware Directory Service (vmdird) and allows an unauthenticated attacker with network access to the target server to bypass authentication protocols. By successfully exploiting this vulnerability, an…

security

Android developer verification will reach every install source, with a side door for power users

Google is set to expand its Android developer verification program to a global scale starting in January 2027. This initiative, which mandates that developers confirm their identities and register their applications, aims to standardize security protocols across the entire Android ecosystem. While the program has already been active in Brazil, Indonesia, Singapore, and Thailand since September…