1300 166 138 Tell us what you need
News / Anthropic

Anthropic opens Mythos cyber AI to Australian defenders

By Monika Tantau • Published 5 months ago
Anthropic Claude logo on smartphone screen, Anthropic Mythos Australia cyber AI

Key takeaways

• Anthropic added Australian organisations to Project Glasswing, granting access to its unreleased Claude Mythos cyber AI model
• The program grew from about 50 mainly American organisations to roughly 200 across 15 new countries, including Australia
• Mythos Preview has reported finding more than 10,000 high or critical severity software flaws across the cohort's targets
• Anthropic confidentially filed for an IPO the day after announcing the expansion

Most Read

Anthropic has expanded Project Glasswing, its gated preview for Claude Mythos, to organisations in 15 new countries including Australia. The Australian Signals Directorate welcomed the inclusion of Australian government and private organisations. Mythos is an unreleased model with strong cyber capabilities, handed to defenders before any public launch.

The expansion took the cohort from around 50 mainly American organisations to roughly 200 globally. For Australian operators, vendors and boards, the practical question is what Mythos can do, who else now has the same access, and how a model gated this tightly gets governed.

What Project Glasswing actually is

Glasswing is Anthropic’s structured preview for Claude Mythos, a cyber-capable model the company has not released publicly. Selected organisations get access under controlled conditions to find software vulnerabilities at scale. Across the cohort to date, Anthropic says Mythos Preview has flagged more than 10,000 high or critical severity flaws.

The June 2026 update widened the program in two ways at once: roughly fourfold growth in organisations with access, and a jump from one main country to 16. Anthropic did not name the Australian organisations brought in. One day after the announcement, the company confidentially filed for an IPO.

Project Glasswing cohort, before and after the June 2026 expansion

Organisations granted access to Claude Mythos Preview. Roughly 4x growth in one round. Source: Information Age (ACS).

Before (early 2026)



~50 orgs
After (June 2026)



~200 orgs

Countries represented also moved from one main country to 16, including Australia.

Why the Signals Directorate welcomed it

The Signals Directorate’s welcome lines up with a defensive logic: putting a cyber-capable model in the hands of national defenders accelerates vulnerability discovery in the same shared codebases attackers probe. Anthropic frames Mythos Preview as a force multiplier for under-resourced security teams, not a new general-purpose model.

The Age framed the announcement as Australia getting access to an AI model some consider too dangerous to release publicly. The tension is real: the same capability that helps defenders find flaws would also help an attacker who could get hold of it. That is the argument for keeping the model gated and identity-bound rather than open.

What it means for Australian critical infrastructure

Operators in power, water, healthcare, communications and hardware are the most exposed surface of this announcement. The pattern is familiar from earlier Glasswing rounds in the United States: vendors that maintain widely reused libraries are pointed at first, defenders work through the findings, and remediation gets coordinated with government cyber agencies. Australian operators may be invited into the same loop.

For boards, the question is no longer whether AI changes the security stack, but how to engage when the offer to participate arrives. Vendor selection, identity binding and a defensible record of model use move from theoretical to operational. Aivy’s guide to choosing an AI automation agency covers the same evaluation discipline more broadly.

Dimension Mythos Preview Typical human red-teaming
Speed across a large codebase Hours to days Weeks to months
Scale of issues surfaced 10,000+ across cohort Tens to hundreds per engagement
Marginal cost per target Low (model time) High (specialist hours)
Access control Gated, identity-bound Contractual, well-understood
Governance maturity Emerging Established practice

Comparison framing based on Anthropic’s reported Glasswing figures (Information Age) and conventional pentesting practice. Marginal cost and governance maturity are qualitative.

The governance question

Wider access to a cyber-capable model creates both defensive value and governance challenges, and the two run on the same axis. Access is gated and identity-bound, and use is reportedly tied to specific defensive scenarios. None of that resolves who is accountable when a finding sits in a vendor’s code that the vendor has not yet patched, or when the same capability is mirrored by a less careful provider next month.

This intersects with policy moves elsewhere, including a US executive order asking for voluntary pre-release testing of cyber-capable models. For Australian boards, that argues for an early position. Our Privacy Act 1988 AI compliance playbook goes deeper into the governance scaffolding this kind of model use needs.

Mythos Preview, reported reach as of June 2026

Three figures Anthropic confirmed at the Glasswing expansion announcement. Source: Information Age (ACS).

Organisations with access



~200
New countries this round



15
High or critical flaws found



10,000+

What happens next

The Signals Directorate’s welcome suggests government and at least some critical infrastructure are early movers. The IPO timing turns Glasswing into more than a defensive program: it is a commercial signal that vendors who can be trusted with cyber-capable models will be a category of their own over the next 2 years.

For Australian businesses outside that circle, the immediate read is supply chain. Your software vendors may now have access to a model that finds flaws in the code you depend on. The right move is asking, not assuming. Aivy’s AI consulting team works with Australian boards drafting the kind of governance position this conversation now demands.

Sources

• Information Age (ACS), “Anthropic gives Mythos AI to Australia’s cyber defenders”. Supports: cohort size (50 to ~200 orgs), 15 new countries, 10,000+ high or critical flaws figure, ASD public welcome, IPO filing timing.

• The Age, “Australia gets access to AI model ‘too dangerous to release'”. Supports: framing of model sensitivity and the defender vs attacker tension cited in section “Why the Signals Directorate welcomed it”.

Scroll to Top