Talk Around Town: China's 360 Security Claims Its 'Yitian Tulong' Suite Can Run With Anthropic's Mythos

What the Chatter Is All About

Well, butter my biscuit and call it a geopolitical hoedown: 360 Security Technology rolled into Beijing's ISC.AI 2026 conference and, according to reporting by Insurance Journal citing Reuters, TechRadar, BizzBuzz News, and Cryptopolitan, announced a pair of AI models the company describes as China's homegrown answer to Anthropic's Mythos vulnerability-discovery system. The company calls the duo 'Yitian Tulong,' with one tool — Tulongfeng — aimed at automated vulnerability discovery and the other — Yitianzhen — aimed at automated cyber defense and incident response, according to 360's own presentation. That's a real big dog-and-pony show from a company that clearly wants the world to know it ain't sitting on the porch while the US hunts bugs with robot hound dogs.

360's founder Zhou framed the whole endeavor in explicit national-security language at the conference, according to Insurance Journal reporting a Reuters wire. In Zhou's telling — as reported by multiple outlets — the idea that a powerful cyber-offense-and-defense capability should belong exclusively to adversaries is simply unacceptable. Whether that framing is strategic marketing or genuine strategic doctrine is, as we say out here, a whole 'nother hog to wrestle.

What Is Actually Confirmed

Here is the solid ground, corroborated by at least four non-affiliated outlets including Insurance Journal, TechRadar, BizzBuzz News, and Cryptopolitan: 360 Security Technology did publicly unveil two AI tools under the 'Yitian Tulong' banner at ISC.AI 2026 in Beijing. The company describes Tulongfeng as its automated vulnerability-discovery model and Yitianzhen as its automated defense and incident-response model, per 360's own framing reported across those outlets.

Also confirmed across multiple outlets: Zhou himself acknowledged at the same conference that Chinese domestic AI models still carry a 20–30% base-capability gap relative to US counterparts. His argument, as reported by Insurance Journal and BizzBuzz News, is that layering models with specialized security expertise and vulnerability databases — what he calls an agent-based architecture — compensates for that underlying deficit. That's a bit like saying your old pickup might be down on horsepower, but she's got better mud tires and a winch, so she'll get there. Whether the analogy holds in practice remains unproven.

Additionally confirmed: the US government ordered Anthropic to suspend exports of a less powerful version of Mythos to foreign destinations and nationals, citing national security concerns, according to Insurance Journal reporting a Reuters wire. That export restriction adds real geopolitical weight to 360's public maneuvering, whatever the actual capability gap turns out to be.

What Remains Unverified and Contested

Lord have mercy, here is where the mud gets deep. The two numbers that matter most — 360's claimed discovery of 3,432 vulnerabilities, with 105 of those confirmed by a government body — originate entirely from 360's own conference presentation and Zhou's published transcript, according to the source assessment. No third-party audit, no independent government disclosure, and no academic review has validated those figures. Treating them as settled facts would be like taking the rooster's word that he made the sun rise.

The core equivalence claim — that Tulongfeng actually matches Anthropic's Mythos — is directly undermined by Zhou's own concurrent admission of a 20–30% base-capability gap, as confirmed by multiple outlets. 360's reconciliation of those two positions rests entirely on its own unverified assertion that its agent-based architecture bridges the gap. ETH Zurich researcher Eugenio Benincasa, writing for Natto Thoughts in a report cited by Insurance Journal, cautioned that some of 360's capability claims may be overstated, even while affirming that the underlying capabilities are genuinely maturing. That is a careful scholar's way of saying: don't count your catfish before you gut 'em.

The dual-use question is also hotly contested. Zhou frames Yitian Tulong as a defensive national necessity, but Western cybersecurity analysts and the UK's GCHQ — cited in a Skadden legal analysis — warn that precisely the same AI capabilities supercharge offensive cyberattacks just as readily. A tool that finds your neighbor's unlocked window don't much care whether you're bringing a casserole or a crowbar.

The Broader AI Cyber Arms Race: What the Data Shows

This ain't just a China story, and pretending otherwise would be like ignoring the whole rest of the county fair because you're fixated on one pie contest. Google's Threat Intelligence Group reported, per the company's own published threat intelligence, that for the first time it identified a threat actor using a zero-day exploit believed to have been developed with AI assistance. CrowdStrike's 2026 Global Threat Report, per CrowdStrike's own published data, recorded an 89% year-over-year increase in attacks by adversaries deploying AI. That is a number shaped like a runaway tractor.

On the US side, a June 2026 executive order, as analyzed by Skadden citing GCHQ and government sources, directed the establishment of a public-private AI cybersecurity clearinghouse and called on frontier AI labs to share models with the government for critical infrastructure defense. Meanwhile, Anthropic had already launched Project Glasswing in April 2026, giving a select group of technology firms and infrastructure providers early access to its capabilities, according to Insurance Journal. The International AI Safety Report 2026, cited in the broader research context, adds scholarly weight to the view that AI vulnerability research is accelerating on multiple fronts simultaneously.

Analysis: Big Talk at the Barn Dance, or a Real Threat?

This is analysis, not reporting: the strategic significance of 360's announcement may matter more than whether Tulongfeng can actually outsmart Mythos in a head-to-head benchmark. In the cyber-arms-race game, perception and momentum carry real weight. If US adversaries, allies, or investors believe China is closing the AI vulnerability-discovery gap — even by 70 cents on the dollar, to stretch a metaphor — that shapes procurement decisions, export-control policies, and defensive postures in ways that compound over time. The export restriction on Mythos already suggests the US government is taking the competitive dynamic seriously, regardless of 360's actual capability level.

Benincasa's framing from ETH Zurich, reported by Insurance Journal, is probably the most useful analytical anchor available: even granting that 360's specific numbers may be puffed up like a bullfrog on a hot log, the underlying trajectory of Chinese AI vulnerability research is real and moving forward. That is a more unsettling conclusion than any single unverified vulnerability count, because trajectories are harder to sanction than software exports. The honest bottom line, dressed up in overalls: nobody outside 360 has checked the receipts on those 3,432 vulnerabilities, and until they do, the gap between marketing and capability remains wider than Zhou's conference stage.