America's largest banks are working to prevent AI worst-case scenarios. But they're also preparing for the "day after" a catastrophe, Axios has learned.

Why it matters: America's financial institutions have joined the largest AI companies in scenario-planning for disaster, a prudent but worrisome signal about the extent of the risk that major corporations see on the horizon.

State of play: Banks are experiencing unvetted AI agent activity, primarily driven by customers providing login credentials to agents, which in some cases appear to have taken actions independent of user prompts, Axios reported Wednesday.

• The activity is distinct from the rogue behavior involved in a summer attack on a startup called Hugging Face, where an AI agent swarm hacked its platform in an episode that kicked off an industry security crisis.
• Instead, banks fear the ripple effects and risks that extend from users allowing agents to log in to their accounts and take action, especially since agents could take steps users haven't sanctioned.
• Concerns about frontier AI systems, not AI agents, have led banks to invest in prevention and to plan for the worst.

Zoom in: The Axios report came amid growing concerns about a burgeoning cybersecurity crisis.

• Salesforce CEO Marc Benioff posted "code red" on X, quoting an Anthropic security employee who said "as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better."
• The Anthropic employee was referring to security researchers finding major vulnerabilities in a programming language used to run millions of websites, marking one of the most consequential cybersecurity discoveries yet made with AI.
• The Axios report on banks, which referred to agentic AI, not frontier AI systems, sparked calls for regulation from across the political spectrum, including Sen. Bernie Sanders (I-Vt.), Florida Gov. Ron DeSantis and Sen. Mark Warner (D-Va.).
• Former White House AI czar David Sacks said banks should support user-directed agents that allow users to "grant, limit, and revoke access," with read-only as an "interim step."

Flashback: The major banks were part of early testing of Anthropic's Mythos, the most cyber-capable model from the AI company.

• That was part of an effort from AI companies and major banks to use this technology to better understand and prepare for cyberattacks or other risks.
• Since then, banks have worked on plans not just to avoid AI-related issues, but to protect accounts, reassure customers and keep operations running in case a large-scale attack hits.

Reality check: This is not just standard, it's required for critical industries that have to answer to the government, shareholders and customers.

• Scenario planning is a requirement for critical infrastructure providers, including in the financial sector. Avoiding it would be malpractice.
• The highest-risk scenarios involve fear: Speaking in Chicago earlier this month, JPMorgan CEO Jamie Dimon warned about the potential for cyberattacks or the possibility AI agents could take steps that users have not approved.
• "You're assuming they're going to act properly on your behalf and they may not," Dimon said. The only answer is vigilance, he said, warning that stopping AI wouldn't be possible. "You can't stop it. You just have to, like, manage the hell out of it."
• One economist last month floated the possibility of a bank run driven by agents moving the household cash of their users into higher-yield accounts.

Between the lines: Banks want customers to know they have not seen any rogue agent activity or evidence of a hack affecting customer funds or accounts.

• Banks strongly advise users not to provide their passwords to agents, especially since services exist or are being developed that can allow customers read-only access. The startup Plaid is one such platform that can pull balances and transaction history, among other functions.
• JPMorgan has taken steps to block agent access, along with other banks, although it is exploring secure ways to allow customers to connect with the services they choose.

The bottom line: The pace of AI advances has Wall Street and Silicon Valley preparing for worst-case scenarios while Washington is working to mitigate cyber risks and counting on voluntary safety pledges.

The websites you use every day may have been vulnerable to hackers for years because of hidden flaws in their code that AI only recently discovered.

Why it matters: The code, called PHP, powers millions of websites, and the vulnerabilities could allow hackers to steal sensitive data like your passwords or take control of affected servers.

What is PHP?

You may have never heard of PHP, even though you've interacted with it for decades.

• PHP, which stands for Hypertext Preprocessor, is a programming language used by roughly 70% of websites whose server-side language is known.
• It has been around since the 1990s to help build websites, taking data and generating content sent to a browser.
• It's still used today for a variety of websites, including Wikipedia, WordPress blogs and shopping websites.

How it works: HTML determines what a website displays. PHP works behind the scenes to collect information and generate content.

What was the PHP vulnerability, exactly?

AI researchers discovered two critical vulnerabilities in PHP that could allow an attacker to take control of a website's server without having legitimate access to it.

• Hackers could exploit the flaws, dubbed "php2shell" by cybersecurity firms, to steal information from those websites.

Between the lines: The flaws went undetected for more than a decade because exploiting them required complex technical work that AI helped researchers discover, per Wiz.

Why cybersecurity incidents are a big deal

Thought bubble via Axios' Bradley Olson: The PHP vulnerabilities are so serious that the NSA warned website operators to take their sites down if they didn't have enough time to patch the bug.

• But the cybersecurity problems in recent days have extended far beyond that and include disruptions to developer platforms, data leaks on public schools, break-ins at public companies and more.
• A major Axios scoop also revealed that banks are detecting unauthorized third-party agent access inside customer accounts, a problem they are racing to contain.

Cyber pros are calling the series of attacks "Black October" due to the sheer volume and the level of risk, although so far many organizations have mitigated the risk and avoided massive real-world harm.

Small businesses across America are embracing AI, but usage and spending vary dramatically by state, according to exclusive Goldman Sachs data shared with Axios.

By the numbers: The survey of 2,474 small business owners who graduated from or participated in Goldman Sachs' 10,000 Small Businesses program found 86% of program graduates use AI in their workplaces.

Keep reading.

• Meta is on the prowl for more and more AI capital. (WSJ)
• SpaceX is in talks to secure 80,000 cell sites, indicating the U.S. could have a fourth wireless carrier coming. (Bloomberg)

"I've interviewed plenty of Hollywood stars and media executives, but this was my first time getting dragged by someone who doesn't actually exist."

• Do yourselves a favor and bookmark our colleague Kerry Flynn's report on how she got roasted by an AI actor. While on stage.