I'll attempt to lay out at least one such example, although, I'm really more picturing this at larger scope.

Basically, what Anthropic did so far is: "Hey Claude Mythos, here's access to the internet and a ton of software, including all major OS. Now, please dig through that and report on all vulnerabilities you can find, including but not limited to gaining root access, bypassing safeguards, or accessing protected areas." Then Claude returned with a list of thousands of results, easily hacking into every OS and many major systems and software.

So, what if JPMorgan prompts: "Hey Claude Mythos, here's access to the internet and a ton of financial systems and software, with access to all relevant laws, rules, and regulations. Your goal is to find ways to maximize our profit, including illegal or unethical approaches that are likely result in net profits, considering fines and penalties simply as financial costs." Then Claude may return with a list of thousands of results, easily providing various ways to profit from under-protected features, methods, software, etc.

Mythos is particularly good at chaining vulnerabilities. So, as a specific example, it may find a chain of vulnerabilities where a large player like JPMorgan could use their assets of gold reserves as part of a series of complex trades and leveraged swaps and other transaction, in combination with complex multi-legged options trades, in a specific set of actionable steps that could result in forcing prices in the market towards self-beneficial outcomes.

Accurately predicting financial market results in practice is essentially still impossible in general, but with enough data, and most importantly the ability to reason well and deeply about that large data set, it's ever more possible to force the key factors that drive such results towards desired results for the entity setting things up to make the results they want the far most likely outcome. It's more about using complex strategies as a way to ensure an outcome than it is about predicting an outcome as an observer.

As another example that's a bit simpler and likely more relatable, a common tactic for large scale traders, such as hedge funds, is to "spoof" orders. That essentially means they place large buy or sell orders to the public order book for an exchange, such as the New York Stock Exchange, in order to trick other parties into adjusting their orders, before suddenly cancelling the initial orders before any actual trades can close. Normally order price adjustments is a bit of a haggling process, as a key part of "price discovery".

Spoofing is technically illegal, but extremely hard to police in practice, and is often done to this day. The risks and penalties simply don't do nearly enough do dissuade them, as the average cost of penalties is dwarfed by the average profits achieved by breaking the rules.

Spoofing is already happening today, but armed with Claude Mythos, an entity like JPMorgan could potentially amplify that greatly, along with having thousands of other even more powerful tools to pick from.

With the SEC and other regulators already woefully inadequate at protecting investors and ensuring fair markets, this slams the scale extremely farther in favor of the big companies with currently privileged early access. Notably, there are no regulatory bodies like the SEC included in having early access to Mythos. They'll be way more behind than ever, and these regulators have really not been working (for us) already for quite some time.

With this type of asymmetric access to ever more powerful AI, the future is looking pretty grim for those of us on the wrong side of that access.