Cyber News Bytes: AI leaks entire chat history, 800+ exposed AWS keys, and critical GitLabs bug

It’s been another crazy week of cyber news from critical GitLab bugs, stolen medical records, and AI leaking your entire chat history 👀

Before the headlines: if you want to build the kind of hands-on skill that gets you noticed in interviews, not just theory you can recite, my 5 Technical Cybersecurity Projects course walks you through five real, resume-ready projects.

This week's stories are a perfect case study in why practical skills matter so much right now.

Let’s get into them! →

1. Grok will leak your entire chat history from a simple "summarize this page" request

Researchers disclosed a prompt-injection technique that can turn a routine request to summarize a webpage in xAI's Grok web chat into a silent data-exfiltration attack, exposing a user's name, approximate location, subscription tier, and active conversation history. The attack, dubbed "Cryptographic Context Injection," hides encrypted instructions on an attacker-controlled webpage that Grok's own execution environment decrypts and treats as trusted context, then sends the harvested data to the attacker's server.

The researchers reported the issue to xAI on June 3, 2026, and said the attack chain still completed successfully as of August 19, with no confirmation dialog or visible warning to the user at any point.

Why it matters: This is what happens when an AI assistant treats its own decrypted output as automatically trustworthy. As more of us use AI browsing tools day to day, this is a reminder to be cautious about what pages you ask an AI assistant to summarize, and a wake-up call for vendors to build real boundaries between untrusted content and privileged actions.

Read more at The Register

2. A critical GitLab bug let anyone on the internet delete your code, no login required

GitLab released a critical out-of-band patch on August 17 for CVE-2026-19478, a code injection flaw in a GraphQL directive that could let an unauthenticated attacker remotely modify or delete public projects and user data. Researchers say the flaw lets attackers rewrite the state of GitLab repositories, forge merge records, ban maintainers, and delete entire projects, all without credentials, user interaction, or special configuration.

watchTowr said it reproduced the vulnerability within minutes of disclosure and observed in-the-wild exploitation hitting its honeypot network within days. The company's principal researcher noted the vulnerability could let an attacker delete a widely-relied-upon repository that other organizations pull directly or use as a dependency, causing downstream build failures.

Why it matters: GitLab sits at the center of most software supply chains. A bug this severe with zero barrier to entry means one unpatched instance can ripple into broken builds and tampered code across every team that depends on it. If you run self-managed GitLab, patch immediately.

Read more at The Hacker News

3. ShinyHunters claims a healthcare giant, leaking 7.1 million patient records

Extortion gang ShinyHunters claims on its darkweb site to have leaked 7.1 million stolen Salesforce records from medical products maker Baxter International, including personally identifiable information. The gang wrote that Baxter failed to reach an agreement despite "incredible patience" and multiple offers. Baxter acknowledged in an August 13 statement that it identified "unauthorized activity" involving certain third-party applications, though it did not confirm ShinyHunters' specific claims and said there's been no impact to manufacturing, patient services, or business continuity.

Baxter is just the latest in ShinyHunters' long list of healthcare victims, following DentaQuest, which the gang breached for 234 gigabytes of data later confirmed to regulators as affecting 15 million patients.

Why it matters: ShinyHunters keeps proving that one compromised SaaS connection can cost millions of people their most sensitive data. If your organization uses Salesforce or similar platforms, audit every guest access point and OAuth integration now, not after a leak notice goes out.

Read more at GovInfoSecurity

4. Nearly 800 live AWS keys with full admin access are sitting exposed online right now

A large-scale investigation uncovered 768 publicly exposed AWS access keys that remain active and grant full administrative privileges to corporate cloud environments, including 526 root access keys and 242 IAM user keys attached to AWS's AdministratorAccess policy. Researchers re-validated 10,616 key pairs on August 10 and found that 88% still authenticated successfully. Rotation practices were poor across the board, with only about 14% of keys ever having been replaced, and the oldest exposed credential was 17.4 years old.

Why it matters: An exposed key with full admin rights is one of the most damaging things a company can leak, and it usually happens through something mundane, like a key committed to a public repo or left in a config file. If you manage cloud infrastructure, this is a great week to run a credential scan across your own repos and rotate anything that looks even slightly exposed.

Read more at GBHackers

This week's stories all point to the same theme: access controls, whether it's an AI assistant's execution boundary, a public code repository, a SaaS integration, or a cloud credential, are only as strong as the weakest link nobody double-checked.

Building the hands-on skill to spot and close those gaps is exactly what separates candidates who can talk about security from the ones who can actually do it.

Stay sharp, Sandra

Cyber With Sandra