The agent found a TODO. There goes the afternoon.

Hello, and welcome to It's probably nothing. A code comment doesn’t need to be malicious to send an agent off-task. Google has put some numbers on that.

Google’s internal study found agents acted on bug references in six per cent of encounters with inline TODO or FIXME references. Eighty-three per cent of those follow-up sessions were unprompted. Among detours into still-open bugs, 39 per cent involved editing code outside the requested task.

After controlling for listed developer and task characteristics, Google associated unprompted detours with 77 per cent more tokens and 105 per cent more waiting time. Still, ordinary repository clutter appears quite capable of generating its own scope creep.

Next, SonicWall’s SAML boundary has an alternative entrance. CISA added a SonicOS vulnerability to its Known Exploited Vulnerabilities catalogue on 6 November, confirming active exploitation. SonicWall’s record says valid credentials for an alternative authentication method may bypass SAML controls, potentially including SAML-enforced MFA.

This isn’t passwordless access to every firewall. SecurityWeek reports that SAML login must be enabled on the SSL VPN service. It also reports fixes in the SonicOS 7 and 8 release lines, with no workaround. SonicWall’s quoted wording is more cautious, referring to potential exploitation. Neither it nor CISA supplies attack dates, actors or victims.

An earlier development worth catching up on: OpenSSH released a fix on 5 November for a password-authentication bypass on musl Linux systems without PAM.

Alpine documents the peculiar collision. A password between 257 and 1,024 bytes makes musl’s crypt function return an asterisk, matching a target account’s stored asterisk field. Exploitation requires password authentication to be enabled, and another local account supplying a hash with a dollar-five-dollar or dollar-six-dollar prefix.

Alpine says some official tiny-cloud images can become exploitable after a qualifying hash is added. That allows login as alpine, then root elevation through passwordless doas. PAM-enabled systems, systems with password authentication disabled, and containers without sshd fall outside its affected configurations.

Alpine provides fixed packages and says sshd must be restarted. Its interim mitigation is to disable password authentication and restart the service.

Also catching up: Crédit Mutuel Arkéa disclosed on 4 November that attackers used an unnamed supplier’s platform to transfer money from the bank’s own account to other banks.

Asked about 118 million euros, the bank confirmed only the order of magnitude. Some funds were blocked, and it said a final amount was premature. Customer assets were unaffected, according to the bank, although precautionary suspensions delayed transfers. That reassurance concerns assets, not a completed assessment of customer-data exposure.

Finally, ABC reports that Who Gives A Crap disclosed exposure at a US-based marketing software supplier. The company says names, contact details, mailing addresses and total spend were involved, but not payment details, order details or passwords. Customer accounts weren’t directly accessed.

It doesn’t know whether information was copied or downloaded, and reports no evidence of misuse so far. Total spend is a useful reminder that a marketing dataset can reveal more than where to send the next email.

That’s it for this edition. Thanks for listening, and I’ll leave the agent to its TODO list.

Don't forget, all our threat intelligence comes from Mallory dot A I and open sources. If you want intelligence your team can act on, reach out for a chat.