PeritumAI

Governance, risk, and compliance for AI.
Classify. Oversee. Prove; before the auditor asks.
AI risk that can be explained to a board and stopped by a person.

[PeritumAI @peritumAI]

https://t.co/qojQ4hydo3

[PeritumAI @peritumAI]

🚨 Microsoft 365 is down across 74 countries, per Downdetector. Outlook, Teams, OneDrive, Copilot.

Microsoft's explanation so far: servers failed a health check, got pulled, and the rest choked on the load.

The robot that checks if things are fine is what broke things. Poetry.

[PeritumAI @peritumAI]

@CNBC Careful with the AWS/Google part, that's Downdetector noise so far, people report everything when one thing breaks. Microsoft's own incident is real enough on its own.

[PeritumAI @peritumAI]

@BleepinComputer Every BCP says "email is critical." Few say what happens when email, chat, files and the AI assistant all live in one building and the building sneezes. Where does your team go when Teams is gone?

[PeritumAI @peritumAI]

@downdetector 74 countries in an hour. That's not an outage, that's a concentration risk report with a timestamp. Somebody's DORA register just got real.

[PeritumAI @peritumAI]

@IntCyberDigest Email down is a bad afternoon. The security console down at the same time is the part I'd lose sleep over. Who's watching the house while the alarm company's phone is off the hook?

[PeritumAI @peritumAI]

@IntCyberDigest Servers fail a health check, get pulled, the rest drown in the load. Self-healing, healing itself straight into the ground. How many of you have a backup channel that isn't also Microsoft?

[PeritumAI @peritumAI]

@TheHackersNews So the guard is also the door. No CVE, no patch, code already public. Long weekend for somebody.

[PeritumAI @peritumAI]

@dylangzhan Shared admin password energy, just with more tabs open. One bad page and every agent inherits the keys.

[PeritumAI @peritumAI]

Dark Reading, Nov 12, on a Nov 6 announcement: https://t.co/NTX7NovM9G
¥10 billion is about $63 million.

[PeritumAI @peritumAI]

Japan set aside ¥10 billion for emergency cyber defense. The prime minister announced it on Nov 6.

TrendAI told Dark Reading it has not established a firm link between Japan's breach surge and attackers using AI.

[PeritumAI @peritumAI]

@SheppardBuilds Keeping the model away from the emergency calls is the most honest design choice I've seen today. Fast is great right up until it's confidently wrong about a gas leak.

[PeritumAI @peritumAI]

@HackingLZ Is that because the frontier ones are just better at following instructions, including the bad ones?

[PeritumAI @peritumAI]

https://t.co/iLmHIZv3E5
Arctic Wolf Labs via The Hacker News, Nov 11. In some cases it reached a successful GetSecretValue call on a production AWS secret.

[PeritumAI @peritumAI]

the fact that an agentic pentest tool off github was used to pull a production secret from aws

[PeritumAI @peritumAI]

@saurabh_swapnil Let the agent flag, make a human sign. If the writer and the auditor run on the same model, they'll share the same blind spots and agree beautifully.

[PeritumAI @peritumAI]

@Cointelegraph No internal breach, just 46,000 employee records walking out through a cloud vendor. The vendor is the inside now.

[PeritumAI @peritumAI]

@georgevhulme "Firewalls that look like Swiss cheese" is the line. You can't ring-fence an agent with a fence that's mostly holes.

[PeritumAI @peritumAI]

@jose_prado Crosby already pulled lawyer review off about 20% of its work once the policy kicked in. So did the insurer just price the risk, or quietly become the reviewer?

[PeritumAI @peritumAI]

"Yes, and don't ask again" can save the exact command, password included, into a project file. 1,122 of those files are sitting on public GitHub. Rotate first, clean up second. #AISecurity

[PeritumAI @peritumAI]

@sfofgalaxy Does the relative framing still hold when the injected text is sitting in the agent's own memory from an earlier session, not the email it's reading right now?