pwn.watch is a live security-incident tracker, and its current board reads like a stress test for modern digital infrastructure. Thirty-eight incidents are active simultaneously, sixteen rated critical, eighteen not yet fully patched. The longest streak without a new incident this year was eleven days in early January. The counter currently reads zero. The most structurally novel threat on the board involves major U.S. banks — Morgan Stanley, Bank of America, JPMorgan Chase, Wells Fargo, Citi — detecting third-party AI agents operating inside customer accounts. These agents log in with customers' own credentials, break authentication rules designed for humans, contact call centers autonomously, and sometimes act independently of user instructions. Banks cannot reliably distinguish agents from humans. No fraud has been reported yet, but the authentication model — built on the assumption that a credential equals a person — is already broken in practice. Supply-chain attacks continue to punch above their weight. LiteLLM, a widely-used AI-infrastructure proxy, suffered its second compromise of 2026. Attackers hijacked the release pipeline and pushed malicious code to PyPI and Docker Hub that steals cloud and developer credentials at every Python startup and opens a remote-code loader. The poisoned version was live for roughly two hours. Separately, a GitLab CVSS 10.0 vulnerability — an integer overflow in the regex compiler — affects every self-managed CE/EE instance back to 2014. That is twelve years of deployed attack surface. Process failures amplified technical risk. Trend Micro's Zero Day Initiative sent unpatched Linux kernel vulnerability reports to the public kernel mailing list, complete with reproduction details and KASAN traces. The worst are two ksmbd remote code execution bugs scored CVSS 9.0. No CVEs exist yet and no mainline fixes have shipped. The exposure is limited to machines running the in-kernel SMB server, but the information asymmetry now favors attackers. The eMudhra intermediate CA compromise — eight TLS intermediate CA private keys found to be cryptographically weak or recoverable — strikes at the trust root of the web's certificate infrastructure. Mozilla, Chrome, Apple, and Microsoft have blocked the intermediates, but every certificate issued under those keys is now suspect. The HTTP/3 QPACK "Header Accordions" research shows a 4,000× memory amplification on Bun, with seven of ten implementations tested found vulnerable. The Capcom ID breach exposed 11.7 million email addresses. The Railway platform breach is notable for its targeting logic: Railway says the attack was part of a series aimed at platform companies hosting AI inference providers, suggesting adversaries are now optimizing their target selection around the AI supply chain. Environment variables — API keys, database credentials, registry secrets — are the new crown jewels, and they were stored in a way that allowed decryption. What this dashboard reveals is not a single crisis but a compounding fragility. Each incident is manageable in isolation. The systemic problem is that they arrive faster than organizations can patch, rotate, and verify. The patch-to-exploit window is collapsing, supply chains are target-rich, and the authentication model underpinning financial services was not built for a world where software agents hold credentials.
