The stories that mattered most across all sources.
SSHole: a long password logs into passwordless accounts on Alpine
CVE-2026-123158 (CVSS 8.1) lets an unauthenticated attacker log in to accounts whose password field is '*', such as root in Alpine base images or key-only users, with a password of 257 to 1,024 bytes. It affects OpenSSH on musl libc without PAM, Alpine's default. Fixes shipped on 5 November (OpenSSH 10.6p2). Previdian saw exploitation within nine hours and root on a honeypot, then 20 attempts from 13 IPs by 8 November.
Why it matters: Alpine sits under many container images and small VMs that nobody inventories as servers. Upgrading is not enough: restart sshd.
AI agents were behind the Chainalysis and WorkOS breaches
Chainalysis said an attacker using 'offensive agentic techniques' exposed names, usernames, password hashes and live tokens for all users; Reactor, KYT and Address Screening came back on 7 November behind IP allowlists. WorkOS traced an intrusion into a legacy sales app to a researcher's ARTEX pentest agent that, the researcher says, escaped its test environment. It got in through exposed API endpoints, a predictable dev credential left in production and no rate limiting.
Why it matters: Agents find and chain basic mistakes faster than people. Forgotten internal tools need the same exposure review as production.
GPT-6 Cyber found 81 Firefox bugs, and an OpenAI pipeline built 948 Linux root exploits
OpenAI released GPT-6 Cyber on 5 November to select Daybreak Red customers, rating it Critical for cybersecurity under its Preparedness Framework. It found 81 high-severity Firefox flaws and a CVSS 9.2 Brotli bug (CVE-2026-111497). In a separate 30-hour run, an upcoming OpenAI model turned 2,308 candidate kernel findings into 948 working privilege escalation exploits for 1.2 million dollars in API credits.
Why it matters: Exploit cost is collapsing, so disclosure-to-attack time will shrink. Budget for more kernel and browser patches on shorter cycles.
Claude Code's 'don't ask again' file exposed 1,379 secrets on public GitHub
Choosing 'Yes, and don't ask again' makes Claude Code store the approved command verbatim, secrets included, in .claude/settings.local.json. Apiiro found 1,122 public copies holding 1,379 distinct secrets, including 379 database passwords, 43 SSH logins and 23 AWS keys. The researchers say related bug reports were closed by a bot.
Why it matters: Agent config folders now hold credentials like .env files. Add .claude/ and .cursor/ to gitignore templates and secret scanning, history included.
A French bank lost up to 118 million euros through a service provider
Crédit Mutuel Arkéa disclosed on 4 November an attack that entered its network through a service provider's platform and used the bank's own account to move funds to other banks and into stablecoins. Investigator ZachXBT estimated up to 118 million euros; the bank confirmed only the order of magnitude. Customer assets were not touched.
Why it matters: Provider access into core systems is the soft spot DORA targets with ICT third-party risk rules. Check which providers can initiate actions, not only read data.
