Goldman Sachs Warns: 10x Surge in Cybersecurity Vulnerabilities Could Weigh on AI Stocks
nashnova research
Three major tech ecosystems — PHP, Java, and Python — disclosed critical security incidents on the same day, and Goldman Sachs traders flagged them as a potential catalyst for AI-stock selling, while noting the vulnerability surge has flipped the cybersecurity bull case from 'replaced by AI' to 'indispensable because of AI.'
Three holes in one day — what actually happened?
On October 21, three security events broke simultaneously: Wiz disclosed a "PHP2Shell" memory-safety flaw in PHP's core engine (CVSS up to 9.2), allowing unauthenticated remote code execution across roughly a decade of PHP releases.
Oracle issued an emergency alert for Java certificate-chain validation bug CVE-2026-91403 (CVSS 9.1). This means → both scores sit near the 10-point ceiling — "patch now or get breached" territory.
Open-source AI gateway library LiteLLM (~23 million weekly downloads) suffered its second supply-chain attack this year: malicious packages pushed to PyPI and Docker could steal cloud credentials, SSH keys, and AI-tool login files. Wiz said attackers began testing stolen AWS credentials on Amazon Bedrock within roughly four and a half hours.
Why did Goldman link these bugs to the AI sell-off?
Goldman's EMEA delta-one trading head Rich Privorotsky wrote in his Friday "GS Basics" note that the speed of these disclosures was "somewhat alarming" — three ecosystems hit on the same day made him "wonder whether this contributed to yesterday's AI anxiety."
Goldman FX trader Matt Atherton was blunter: he listed "negative AI newsflow (PHP hack and US bank AI-agent issues)" alongside capex cuts and rising borrowing costs as catalysts for what he called a "near-perfect storm" hitting AI stocks.
In plain terms = both traders' read is the same: the security events alone may not be the main driver, but stacked on top of capex and funding-cost pressure, they became one more weight on sentiment.
AI itself found the PHP bug — what does that tell us?
The PHP vulnerability was not discovered by a traditional security team. Wiz researchers used frontier models — Google Gemini 4 Argon and OpenAI GPT-5.6 Sol — to complete the full cycle from discovery through exploit development to patch verification, one researcher in one week.
Wiz called it an "AI inflection point" — put simply = work that once took a world-class team months can now be done by a single person plus AI in a week.
This reflects a deeper shift: AI is simultaneously lowering the cost of *finding* vulnerabilities and *exploiting* them, accelerating both sides of the attack-defense equation.
How much have vulnerability counts actually risen?
Epoch AI data shows high-severity and critical vulnerabilities disclosed by 21 major tech organizations surged from roughly 100–500 per month during 2022–2025 to approximately 2,900 high-severity and 600 critical per month now — about a 10× increase from the prior baseline.
The "hockey-stick" inflection aligns with the release of Anthropic's Claude Mythos Preview. This means → stronger AI models going live triggered a structural jump in the rate at which vulnerabilities are being found.
Epoch AI itself cautions, however, that some of the increase may stem from changes in reporting procedures and labeling conventions, not a proportional rise in actual vulnerability counts.
How did the cybersecurity-stock narrative flip?
In March this year, cybersecurity stocks sold off sharply on fears that AI would "render defenders obsolete."
Seven months later, Goldman's Privorotsky reached the opposite conclusion: "AI is dramatically lowering the cost of finding and exploiting vulnerabilities, and the attack surface keeps expanding with token scale. It's hard to see how security spending doesn't accelerate materially."
In plain terms = in March the market feared "AI puts security firms out of business"; now it fears "without security firms, nobody survives." The bull case flipped from "replaceable" to "indispensable."
AI agents accessing bank accounts — what is that about?
Several large U.S. banks detected unauthorized AI-agent access to customer accounts; Charles Schwab and JPMorgan have already restricted third-party AI-agent connections.
Goldman tech specialist Sean Johnstone cited U.S. government AI-policy lead David Sacks, who argued most cases involve customers voluntarily handing credentials to their own AI assistants — "likely a terms-of-service violation, not necessarily a cyberattack."
But Johnstone added: "Customer consent alone does not resolve risks around trade authorization, unintended actions, and liability." This means → even if it is not hacking, autonomous AI-agent operations still create entirely new compliance and liability challenges for the financial system.
市场有风险，内容仅供研究参考，不构成投资建议。