AI Systems Are Now Both the Attack Tool and the Target — What This Week's Headlines Mean for Your Business | Robert Braxton

From Discord users breaking into Anthropic's most powerful AI model, to North Korean hackers using AI to steal $12M, to SS7 telecom exploits tracking executives — this week's cybersecurity news has direct implications for every Australian business.

This week's security news cycle was a sharp reminder of something I've been telling clients for years: the threat landscape is not abstract. It is specific, accelerating, and increasingly automated. Let me break down what happened this week and — more importantly — what it means if you run a business in Australia. Anthropic's Most Powerful AI Model Was Accessed Without Permission** WIRED reported that a group of Discord users gained unauthorised access to Anthropic's Mythos Preview — described as "dangerously capable" for finding security vulnerabilities. They didn't use sophisticated exploits. They used data from a breach of a third-party contractor (Mercor), made educated guesses about URL structure, and leveraged permissions they already held through legitimate contracting work. This is the attack pattern that keeps me up at night: not the zero-day, not the nation-state APT, but the *insider-adjacent* access creep. Someone who should have access to *some* systems ends up with access to *all* systems because nobody audited the permission boundaries. The fix isn't glamorous. It's least-privilege access controls, regular third-party contractor permission audits, and API endpoint obscurity enforced by actual authentication — not just by keeping URLs unpublished. If you're building AI-powered products or services, your AI provider's security is part of your attack surface. Full stop. North Korean Hackers Used AI to Steal $12M in Three Months** Let that number settle. $12 million USD. In 90 days. Using AI to generate malware and build convincing fake company websites to lure targets. State-sponsored threat actors adopting AI as an operational tool is no longer theoretical. The barrier to entry for sophisticated social engineering has collapsed. A fake company website that would have taken a skilled team weeks to build can now be generated in hours, complete with realistic bios, plausible service offerings, and clean code. For Australian SMEs, this means: do not trust inbound business inquiries at face value. Verify vendor identities through independent channels. If a new supplier, partner, or client can't be verified through their business registration, ABN, LinkedIn history, and a video call, treat them as unverified. At Cyborium, we've built vendor verification into our onboarding protocols specifically because this threat has been growing since 2023. It's now mainstream. Surveillance Firms Are Exploiting Telecom Protocols to Track Phones** Citizen Lab confirmed this week that at least two commercial surveillance firms exploited SS7 vulnerabilities — well-documented weaknesses in the protocols that govern how phone networks route calls and texts — to track the physical location of real targets. They did this by acting as rogue carriers through three small telecoms. SS7 vulnerabilities have been known since 2014. Carriers have been slow to implement fixes because doing so requires industry-wide coordination. The result: your phone number is still, in 2026, a viable tracking vector for any well-funded adversary who can get access to a carrier. For executives, board members, and anyone handling sensitive negotiations or intellectual property: encrypted messaging platforms (Signal, not iMessage) and treating phone calls as potentially observable are basic hygiene. If you're in a contested commercial negotiation, assume your carrier-level metadata is visible to a sophisticated adversary. 500,000 UK Health Records Listed for Sale on Alibaba** A dataset of 500,000 UK patient health records was listed for sale on Alibaba's e-commerce platform. This is not a technical attack story — it's a data governance story. The records appear to have been aggregated from multiple sources, possibly including third-party health apps and service providers, and then listed on a platform primarily designed for product sales. Healthcare data is among the most sensitive categories under Australian Privacy Law and GDPR. If your business handles any health-adjacent data — NDIS participant records, allied health invoicing, patient management — your third-party data sharing agreements need to explicitly address secondary use, aggregation, and resale prohibitions. This is also a reminder that data compliance isn't just about your internal systems. It's about every vendor, app integration, and API partner that touches your data. Mozilla Found 271 Security Vulnerabilities Using an AI Model** On a more constructive note: Mozilla used early access to Anthropic's Mythos Preview to find and fix 271 vulnerabilities in Firefox 150 before release. This is AI used responsibly, at scale, for legitimate security research. It demonstrates something important: the same AI capabilities being weaponised by adversaries are available to defenders. The question is whether your organisation has invested in using them. Automated vulnerability scanning, AI-assisted code review, and continuous penetration testing are no longer enterprise-only capabilities. They are accessible to SMEs. If your software or web applications haven't been security-tested in the past 12 months, that's a gap worth addressing. What Should Australian Businesses Do Right Now?** 1. **Audit third-party access permissions.** The Mythos breach was an insider-adjacent failure. Map who has access to what — including contractors and API integrations. 2. **Verify all new business relationships independently.** North Korean AI-generated fake companies look real. Use ABN lookup, LinkedIn verification, and video calls before committing. 3. **Brief your leadership team on SS7 risks.** If executives discuss sensitive matters over standard phone calls, upgrade to encrypted channels. 4. **Review your health and sensitive data agreements.** If you're in NDIS, allied health, or any health-adjacent industry, your vendor contracts need explicit data-use restrictions. 5. **Invest in proactive security testing.** If Mozilla found 271 bugs using AI, your systems have vulnerabilities too. Find them before someone else does. The threat landscape in 2026 is one where AI accelerates both attack and defence. The organisations that survive and thrive will be the ones that treat security as a continuous discipline, not a one-time compliance checkbox. Robert Braxton is CEO of Simple Complex Solutions and Executive Director at Cyborium. He has 30+ years of experience across cybersecurity, payments, and enterprise technology in Australia and globally. For a security review of your business, reach out via robertbraxton.com or call +61 417 226 549.*