At Menagos, we deliver intelligent, adaptive cybersecurity solutions designed to stay ahead of modern threats. Our AI-driven monitoring, and expert consulting.
Fifth: set cost and loop limits before you need them. An agent retrying a failing step will burn budget all night without complaint. OWASP lists this as LLM06 Unbounded Consumption. Cap calls, time and spend per task. #AISecurity 5/5
We run multi-agent AI operations every day. Five things that broke, and what we changed.
First: broad agents drift. One agent that could research, write and act did all three worse. We split work into narrow roles, each with only the tools that role needs. #AgenticAI 1/5
Fourth: log every tool call or you cannot debug. The final answer tells you something went wrong, not where. We log each call, its inputs and its result, so a bad outcome traces back to one step. 4/5
Freshness depends on the artifact. A policy, a log export, a scan report and a training record age differently. One question before any assessment: for each control, what is the newest artifact, and has anyone checked it against actual practice? 2/2
Your access control policy was reviewed 18 months ago. Is that evidence or history? CMMC readiness rarely stalls on control mapping. It stalls when someone asks where the evidence is, and whether it is recent enough for an assessor to accept. #CMMC 1/2
Caveats: npm only for now, Enterprise Cloud only, and it only protects traffic routed through it. Still the right shape. Give agents least privilege, no standing secrets, and a gate on what they can install. https://t.co/6ZuH3ZsooI 3/3
Your AI coding agent runs npm install, often with credentials sitting in its environment. Install scripts run before any test or scanner looks. That is the AI supply chain problem in one sentence. #AISecurity 1/3
GitHub put a package firewall in public preview on Oct 20. It sits between registries and CI runners, dev machines and AI agent sessions. Policies cover malware advisories, vulnerable versions, attestation downgrades, newly added install scripts and package age. 2/3
CISA's KEV due date for two Plex2Shell CVEs (CVE-2026-96654, CVE-2026-96656) is today. The other two, 96652 and 96655, are due Nov 3. Fix is Plex Media Server 1.43.3.10861 or later. If Plex runs on a box near work data, check it. https://t.co/nVvpKxmDqM
So the defender math changes. NorForge found successful break-ins took just under 23 hours on average, at roughly $5 a task. Patch speed and web tier telemetry now decide who gets hit. https://t.co/KzwDBNpUxy 3/3
Someone pointed autonomous AI agents at 1,200+ websites and got into sites belonging to 100+ companies, per NorForge's research and Reuters' review. The operator's instructions were under 100 Chinese characters. The agents did the rest. 1/3
Nothing in the chains was exotic. SQL injection, unsafe uploads, deserialization, auth bypass, then a web shell uploader for persistence. AI didn't invent new bugs. It made old bugs cheap to find at scale. 2/3
Not everyone online is who they claim to be! Phishing attacks trick you by pretending to be trusted people or companies — just to steal your info. Stay alert, double-check emails, and never click suspicious links!
#PhishingAwareness#CyberSecurity#OnlineSafety#DataProtection