On 7 April 2026, Anthropic announced a frontier AI model called Claude Mythos Preview. Within a week, it was the only thing anyone in our industry wanted to talk about. Clients emailed. Our consultants debated it in Slack. Conference organisers started rewriting agendas. And the question behind all of it was the same one we have been sitting with ourselves: what does this actually mean for the way we do penetration testing?
This post is our honest attempt at an answer. Not the marketing version. The version we would give a CISO over coffee.
What Mythos actually is
Mythos is not a pentesting product. It is a general purpose reasoning and coding model that happens to be unusually good at the work that sits at the heart of offensive security. Reading unfamiliar code. Reasoning about memory safety and control flow. Chaining small weaknesses into working exploits. Doing all of that at machine speed.
Anthropic decided not to release it publicly. Instead they set up something called Project Glasswing, a coalition of big tech partners and critical infrastructure operators who will use the model to find and fix flaws in foundational software. In the weeks before the announcement, Anthropic used Mythos internally to identify thousands of zero day vulnerabilities across every major operating system and every major web browser, including bugs that had sat untouched for 16 and 27 years despite years of review.
Whether a consultancy like Intrix ever gets direct access to Mythos itself is an open question. But the capabilities it demonstrates will be in production models generally available to the market soon enough. That is what we are planning for, and what we are advising our clients to plan for.
Where this genuinely makes us faster and better
A good penetration test is three phases stitched together. Reconnaissance and surface mapping. Vulnerability discovery and exploitation. And the human work of interpretation, impact analysis, and reporting. AI of this calibre compresses the middle phase dramatically, and that changes the economics of everything around it.
The most immediate gain is coverage. A typical external web application test runs for a fixed number of days. Inside that window our consultants make judgement calls about where to spend time. Which endpoints to fuzz. Which authentication flow to pick apart. Which dependency to audit. AI assisted discovery lets us parallelise that work. While our lead consultant focuses on business logic abuse and privilege escalation paths that still need human reasoning, an agent can be grinding through dependency chains and subtle flaws in the background. The engagement does not get shorter. It gets denser.
Whitebox testing also gets a lot more valuable. Some of the most useful real world data we have seen on this comes from Aikido, who reported that whitebox tests with source code access surfaced roughly seven times more critical and high severity issues than greybox tests, and ran at roughly twice the efficiency. That matches what we have been seeing on our own engagements. Clients who historically chose blackbox testing because it felt more like a real attacker are now in a position where the economics tip decisively towards giving testers source code. More context, more findings, faster.
Then there is the long dormant bug problem. The Mythos results included flaws that had been sitting in widely used open source code for over a quarter of a century. That class of vulnerability, deep, non obvious, requiring sustained reasoning across a large codebase, is exactly what human pentesters struggle to find inside a commercially viable budget. AI assistance meaningfully changes what is economically discoverable, which means our clients can now afford to look for things they previously could not.
Finally, there is the shift away from annual tick box testing. A once a year pentest was always a compromise with reality. With AI compressing discovery cost, continuous and near continuous assurance becomes affordable for mid market Australian clients, not just ASX 100 security teams. This is where our vCISO and managed services work starts to merge with our pentesting practice. The deliverable stops being a PDF that lands in April and becomes an always on posture.
The honest downsides
We are not going to pretend this is a clean win. Any consultancy telling you it is a clean win is selling you something.
The most obvious problem is that attackers get the same capability curve. Every efficiency argument we made above applies just as cleanly on the other side of the fence. Mythos and models like it could change the basic economics of cybersecurity. Serious vulnerabilities have often stayed hidden because finding them took rare skill, patience, and time. If these models can scan the hidden plumbing of the internet at unprecedented scale, what is now specialised hacking becomes a routine, automated process. Anthropic’s restricted access strategy buys time, not permanent safety. Open weight models are closing the gap, and the window where defenders have a genuine edge is narrow.
The less obvious problem, and the one the industry is not talking about loudly enough, is remediation. If AI finds ten times as many real vulnerabilities, who is fixing them? Patch management, change control, regression testing, and developer capacity do not scale with model capability. Forrester put it bluntly: discovery is now cheap, remediation is where the value lives. Finding things is easy, fixing them is hard. For our clients, that means a pentest report with 200 real findings is useless without prioritisation, sequencing, and remediation support. The deliverable has to evolve. Raw vulnerability lists are about to be table stakes, and the real work is in ruthless, context aware triage.
Pricing is going to get uncomfortable for parts of our industry. Traditional penetration tests have historically priced off the perceived scarcity of discovery expertise. If discovery stops being scarce, that pricing model comes under pressure. Firms that sold box checking pentests to satisfy compliance will feel this first. Firms that sold judgement, context, and strategic outcomes will feel it less. We know which side of that line we need to be on.
There is also a subtler issue that the AISLE research team put forward as a useful counterweight to the Mythos hype. They showed that cybersecurity capability is jagged, that it does not scale smoothly with model size or price, and that smaller cheaper models recovered much of the same analysis once the relevant code path was isolated. The practical lesson is that raw model capability is not the moat. The moat is the scaffolding around the model. The orchestration layer. The security domain context. The judgement about what to test, how, and in what order. A consultancy that wires up the latest model and slaps an AI label on itself will be outperformed by one that builds deep security expertise into the tooling.
What this means for graduates entering cyber
This part matters to us. Intrix is active in the Australian cyber community, we hire juniors, and we get asked by university students at every BrisSec and every career day whether they are entering the industry at the wrong time. The honest answer is complicated, so here is the long version.
The bad news first. A lot of what used to be junior offensive security work is going to be done by an agent. Running scanners, triaging their output, writing up low hanging findings, doing the first pass on a well understood web application. If you were expecting to spend your first two years grinding through that kind of work to build skills, that path is narrowing. Firms will need fewer people to do it, and the people they do need will be expected to oversee an AI doing it rather than doing it by hand.
The good news is that the skills that actually matter on day one of a career have not changed. They have become more important. Understanding how operating systems actually work. Being able to read source code in a language you have never seen before and reason about it. Understanding networks at the packet level, not the dashboard level. Being able to sit with a problem for three hours and not give up. Writing clearly enough that a non technical executive understands what you found and why it matters. Those are the foundations, and they are exactly what separates someone who can supervise an AI agent usefully from someone who cannot.
Our advice to graduates and career changers, as a firm that is hiring, is this. Do not try to compete with the model on things the model is good at. Learn the underlying systems deeply enough that you can tell when the model is wrong, because it will be wrong, often in ways that look confident. Build a portfolio of work that shows judgement, not just output. A write up of a bug you found and the reasoning you used is worth more to us than a screenshot of a scanner finding. Learn to communicate. A junior consultant who can run a client workshop and translate technical risk into business language is more useful to us in year two than one who can run fifty tools but cannot sit in front of a board.
Also, and this is important, do not skip the fundamentals because the AI can do them. We have a rule internally that our consultants need to be able to do the work without the AI before we let them do it with the AI. That is not a hazing ritual. It is how you end up with people who can actually supervise the tooling, catch its mistakes, and push back when a model produces a confident looking finding that is nonsense. If you are studying now, the best thing you can do is the same thing the best pentesters have always done. Break things, read the source, build the mental model from the ground up.
The last thing we would say to graduates is that the ceiling for a good cyber career has probably gone up, not down. The work that AI cannot do, deep adversarial thinking, business context, remediation partnership, running a red team engagement against a sophisticated defender, is going to be in more demand and better paid, not less. The floor has risen too. You cannot coast on knowing how to run a scanner. But for someone willing to do the real work, this is a genuinely good time to be starting.
How Intrix is actually adapting
We would rather tell you what we are doing than issue a vague statement about being AI ready. So here it is.
We are integrating AI assistance into our testing workflow methodically, in specific phases where we can measure the quality of the output and catch hallucinations. Dependency analysis. Code review support. Attack surface enumeration. Payload generation. We are not replacing the consultant led kill chain. We are making the consultant faster and more thorough.
We are also reshaping our deliverables. Our reports already emphasise business impact over raw CVSS scores, and we are going further. Every finding now carries explicit remediation sequencing, and our retest and continuous assurance offerings are being expanded so clients can actually close the loop at the pace AI accelerated discovery is going to demand.
And we are investing in the tooling layer. Through our sister venture Attack Insights, we are building autonomous penetration testing infrastructure that brings AI driven discovery under scaffolding designed by working pentesters. The goal is not to replace the consulting engagement. It is to give our consultants and our clients a shared platform for continuous validation between formal tests.
The bottom line
Mythos is the clearest signal yet that the economics of penetration testing are changing. Discovery will get cheaper, faster, and more comprehensive. That is genuinely good for Australian businesses who have historically been under tested. Small and mid market firms that could never afford meaningful offensive assurance are about to be able to. It is less good for consultancies whose business model depended on discovery being scarce, and it is going to be a confusing few years for people entering the industry.
Our read is simple. The firms that will do well over the next three to five years are the ones who treat AI as a force multiplier for skilled humans, who invest in the orchestration and judgement layer rather than chasing the latest model, and who reorient their deliverables around the work that still genuinely requires a person. Prioritisation. Business context. Remediation partnership. Adversarial creativity that does not yet live in a language model.
That is the firm we are building. If you want to talk about what any of this means for your own security programme, whether you are running annual tests today and wondering if that is still enough, a CISO trying to brief your board on AI era risk, or a graduate trying to work out where to point your career, we are always happy to have the conversation.
Intrix Cyber Security is a CREST accredited Australian cybersecurity consultancy providing penetration testing, red teaming, GRC, vCISO, and managed security services. Get in touch at intrix.com.au.
