In-depth security research across CVEs, incidents, TTPs and defensive practice. When a post is rule-bound, the canonical CloudSigma rule evidence is named in the article.
A passkey assertion can be cryptographically valid even when no genuine user verification occurred or device-enrolment trust was abused; correlate endpoint evidence, the WebAuthn ceremony and a deliberately linked cloud session.
TA488's Zimbra and Outlook Web Access campaigns show why patching and password resets do not finish a webmail incident; defenders need to hunt the mailbox ACL, application passcodes, browser storage and outbound traffic left behind.
CVE-2026-54121 can turn a low-privileged AD CS request into Domain Controller identity. Patching stops new abuse, but only a cross-layer hunt can show whether the domain was already compromised.
GitHub now delays routine Dependabot updates for three days while PyPI closes release file sets after 14, leaving a detection gap that release-age and file-timestamp joins can expose.
Laundry Bear's Zimbra campaign leaves a behavioural chain across SOAP activity, mailbox archive requests and DNS that defenders can hunt after its infrastructure changes.
The AsyncAPI compromise shows why disabling npm install scripts is not enough: the malicious loader waited for first import, then spawned a detached process with persistence and command-and-control.
ShinyHunters-linked Salesforce Experience Cloud abuse is a public-portal configuration problem, but defenders should treat suspicious Aura queries as the start of CRM data theft and follow-on voice phishing.
A fixed Vertex AI SDK bucket-squatting flaw created a narrow path to model poisoning and credential theft. Detection depends on joining SDK versions, bucket ownership, deployment events and runtime token use.
ShinyHunters exploited exposed Oracle PeopleSoft management paths before deploying remote-access tooling and moving towards extortion. Detection depends on separating victim-side PSEMHUB evidence from attacker staging activity, then joining edge, host and egress telemetry.
Langflow exploitation is not only a patch race: exposed AI workflow servers can turn one public-flow RCE into environment-secret theft and cloud token use.
FortiBleed forces cloud defenders to join VPN, identity and cloud-audit telemetry around recovered credentials, not stop at Fortinet patch status.
Cloud log changes should not be triaged as isolated admin events when they sit between valid-account access and cloud data theft.
Analysis of the KnowledgeDeliver CVE-2026-5426 exploitation chain, moving from unauthenticated ViewState deserialisation via hard-coded keys to in-memory Godzilla web shells.
SaaS vishing incidents now move from helpdesk pressure to SSO abuse and repository export fast enough that defenders need one identity-to-SaaS detection chain.
Kubernetes service-account token theft now needs cloud-control-plane detection, because the meaningful loss often happens after the pod boundary is crossed.
Browser and research agents need detection around the tool boundary: canary tools, planted credentials and cloud audit joins catch prompt injection when prevention misses.
Flowise and Semantic Kernel flaws show the same defensive problem: a prompt event is weak evidence until it is tied to process execution, secret access, and cloud API use from the workload identity.
Microsoft, GMO Flatt and Anthropic now show the same weak point in agentic CI: untrusted repository text can steer a coding agent towards runner secrets unless defenders bind GitHub events to cloud token use.
Unit 42's autonomous cloud-agent test shows the useful defender signal: AI-driven attacks still leave cloud audit traces when they enumerate, impersonate and exfiltrate.
Prompt injection is becoming an operations problem: defenders need to join agent prompts, tool calls, SaaS access, and cloud audit logs before a hidden instruction turns into data access.
MCP security research is strong on attack design and thin on runtime detection. Here is a reproducible model that maps tool poisoning, output poisoning, sampling abuse and covert tool invocation to concrete signals in MCP and agent audit logs.
The current TanStack and Mini Shai-Hulud wave is less about poisoned packages and more about stealing the developer tokens that let attackers publish the next wave.
Unit 42's AD CS research is a useful reminder that certificate services are now part of the identity attack surface, not a quiet infrastructure corner.
Canonicalized CloudSigma rule for CVE-2026-42823 (T1548). Improper access control in Azure Logic Apps allows an authorized attacker to elevate privileges over a network.
Google says it found an AI-assisted zero-day exploit, but the harder problem is the AI software supply chain now becoming an initial access path.
MOVEit Automation CVE-2026-4670 is a critical authentication bypass in managed file transfer automation. Patch with the full installer, then investigate access to transfer jobs and stored credentials.
GitHub Enterprise Server CVE-2026-3854 turns repository push access into possible server compromise. Patch GHES, review push logs, and check CI trust paths.
cPanel and WHM CVE-2026-41940 is a KEV-listed authentication bypass in a hosting control plane. Patch the service, then investigate sessions and tenant impact.
Reviewed CloudSigma rule for CVE-2026-31431 (T1068). In the Linux kernel, crypto algif_aead reverted to operating out-of-place.
LiteLLM CVE-2026-42208 turns an exposed model gateway into a credential exposure problem. Patch first, then treat stored provider keys as potentially read.
CVE-2023-34192 is a stored XSS flaw in Zimbra Collaboration Suite. The CloudSigma rule detects CRS 941xxx XSS blocks against the autoSaveDraft endpoint.