The Warning 100+ Companies Signed

The letter's warning about unaccountable AI already has a case study: while most security programs spend their AI energy policing who pastes company text into ChatGPT, Akamai's research points somewhere else. The top 5% of enterprise power users are hardwiring unvetted AI tools into production workflows, customer-facing systems, and data pipelines. That is not casual usage. It is unreviewed integration surface, built by the most capable people on the team, sitting in the paths that matter most.

The controls meant to catch this are shakier than the marketing suggests. Researchers this week showed a Cryptographic Context Injection technique that slips past safety guardrails in Grok and Gemini by hiding instructions until they decrypt inside the model. The tl;dr sec shadow-AI briefing puts the through-line plainly, warning that a shadow app integration today is a supply-chain attack vector tomorrow. A helper browser extension or an AI connector wired into a build is a dependency nobody added to the SBOM.

Red Hat shipped an emergency patch for a Keycloak flaw rated CVSS 9.1 that lets an unauthenticated attacker take over any account by forcing a password reset. In the same window, GitLab CVE-2026-19478 came under active exploitation within days of disclosure, and CISA added actively exploited Zimbra and TrueConf Server flaws to its Known Exploited Vulnerabilities catalog, the latter already used by the Head Mare group to deploy PhantomCore malware.

The common thread is timing. The gap between a public advisory and a working exploit keeps shrinking, and it now runs shorter than most enterprise change windows. When the identity platform itself is the vulnerable component, there is no upstream boundary to hide behind while you schedule the fix. The patch and the exploit are now racing on the same clock, and the exploit had a head start this week.

Anthropic expanded Claude Security's Mythos 5 codebase scanning to more enterprise customers and announced a $35M open-source security fund. It arrived alongside GitHub's own scorecard: its Secure Open Source Fund reported $500,000 across 50 projects in one session, and $1.88M across 188 projects to date. Two of the companies whose models generate the most new code are now directly funding the maintainers that code leans on.

The timing is not charity theater. This same period, attackers with North Korean infrastructure overlap compromised the arrayref Rust crate and slipped in a dependency that ran a payload at build time, before any review. GitHub's work with 50 projects named the limit of the money plainly, finding that AI helps maintainers triage faster while they still supply the judgment and accountability for what ships. The funding buys tooling and time; it does not buy the human review that catches a poisoned crate.