Last week, Anthropic introduced Claude Mythos Preview and Project Glasswing. Anthropic says the model has identified thousands of previously unknown vulnerabilities across critical software, including flaws that earlier-generation models missed.
The cybersecurity industry reacted immediately. Shares in CrowdStrike, Palo Alto Networks, Zscaler, and SentinelOne dropped between five and eleven percent in a single session. Not because of a breach. Because of what the announcement implied: AI can now find and exploit vulnerabilities faster than humans can patch them.
Apple, Microsoft, Amazon, Google, Nvidia, and JPMorgan Chase all signed on to Anthropic’s emergency Project Glasswing coalition, a coordinated effort to patch the most critical systems before Mythos-class capabilities become widely available. That list of names tells you everything you need to know about how seriously this is being taken.
There will be more attacks, faster attacks, and more sophisticated attacks. Now is the time to modernise cybersecurity stacks everywhere.
The technical threat is real. But it is not your only problem.
Here is the number that does not appear in the Mythos coverage, but should: 68% of all data breaches involve a human element. Not a zero-day vulnerability. Not a sophisticated exploit chain. A person clicking a link, forwarding the wrong document, approving a payment that arrived at exactly the right moment with exactly the right context.
Attackers do not need Mythos-class AI to compromise your organisation. They need your CFO to receive an email that looks like it came from your CEO, referencing a real deal, using real supplier names, and asking to move money before Friday. That email is getting easier to write, not harder. The same AI capabilities that Anthropic is using defensively are available, in various forms, to the people on the other side.
CrowdStrike’s 2026 Global Threat Report found an 89% year-on-year increase in attacks by AI-enabled adversaries. The important point is not that every attacker suddenly has the same capability. It is that reconnaissance, impersonation, and attack preparation are becoming faster and easier to scale.
And yet. The training has not changed.
Most organisations running security awareness training today are running the same programme they were running last year. The same generic module. The same compliance checkbox. The same video your employees click through in four minutes while making coffee, forget by lunch, and pass the quiz on because the questions have not changed in three years.
Mimecast’s State of Human Risk 2026 report found that just 8% of employees account for 80% of all cybersecurity incidents. Eight percent. The training that treats every employee identically is missing the people who actually need it most, and doing very little for everyone else.
The threat landscape just changed overnight. The training budget meeting is next quarter. This is the gap that attackers live in.
- 68%
- of breaches involved a non-malicious human elementVerizon DBIR 2024
- 89%
- year-on-year increase in attacks using AI by adversariesCrowdStrike Global Threat Report 2026
- $13.1M
- average estimated cost per insider-driven incident in 2026Mimecast State of Human Risk 2026
The problem was never awareness. Every CFO who has ever clicked a fraudulent link knew what phishing was. The problem is relevance, and the industry has spent two decades solving the wrong thing.
What actually changes behaviour.
I worked as an incident response analyst at one of the UK’s largest telecoms providers. I watched a CFO fall for a targeted phishing attack after completing mandatory annual training. The training had told him what phishing was. It had not shown him what it looked like when it arrived in his specific inbox, from a sender who knew his name, his role, and his supplier relationships.
The research on this is consistent and uncomfortable. Knowing that something exists does not protect you from it. Relevance does. Training that references your world, your role, your near-misses, your specific workflows, changes behaviour in a way that generic content simply cannot.
This is what Excera is built around. Safe social-engineering simulations shaped by the person’s role, authority, language, and working context. The point is not another lesson. It is to test the decision an attacker would try to bend, then show the person what gave the pretext away.
The session did not exist before the person logged in. Ten minutes. No passive slides. Three specific behaviours to implement before Friday, calibrated to their actual authority level and workflow.
The window is narrowing.
Anthropic is restricting Mythos Preview to a gated research programme for defensive use. That restraint does not change the direction of travel. AI-assisted vulnerability research and AI-assisted deception will keep improving.
The organisations that will weather that shift are not the ones with the most sophisticated firewall. They are the ones whose people know what a targeted attack looks like when it arrives in their specific inbox, from a sender who knows their name.
The technical defences are being upgraded. Is the human layer?