Microsoft FORGE Lab researchers worked with @github and the maintainers of six open-source projects to test whether AI-assisted vulnerability discovery can deliver verified, fixable reports instead of raw findings. msft.it/6010xEdcp
Codename MDASH, using Microsoft's MAI Cyber model, flagged thousands of possible issues. After triage and manual review, maintainers received 99 private reports, each with a proof of concept and a suggested fix.
The pipeline had three stages:
-MDASH scanned each project and proposed fixes.
-An AI triage agent from GitHub Security Lab removed duplicates, checked validity, and prioritized high-impact bugs.
-Seven FORGE Lab researchers then validated remaining findings by hand, spending about two days per project.
The projects were AutoGPT, Caddy, Express, FastMCP, ImageMagick, and Netty. Reports went through GitHub private vulnerability reporting, and maintainers decided what to fix and when. The first fixes shipped about a week after the initial scan.
The research also shows the limits. Manual validation doesn't scale yet, and filtering wasn't perfect: at least one valid report duplicated an issue maintainers already knew about. Each lesson is an opportunity. GitHub and Microsoft will keep working together to help secure open source.
Read the full write-up for the methodology, maintainer feedback, and what's still needed to make AI-assisted vulnerability research work for open source.
Microsoft research shows how device code phishing, which persuades users to approve a legitimate sign-in rather than stealing a password, can extend to AI coding agents, using GitHub Copilot CLI as an example. msft.it/6012xEPT6
In a controlled proof-of-concept simulation, one approved device code led to four possible outcomes: remote steering of an active session, misleading entries in agent memory, exposure of synced session history, and access to private repositories. None of which were observed in the wild.
Any device-flow client can be targeted this way, and a legitimate verification page does not mean the request is safe. Microsoft Defender detections and disruption help connect phishing attempts to GitHub authorization activity. Read the blog for the full findings and mitigation guidance.
Microsoft Threat Intelligence has published a detailed analysis of this phishing campaign. Read the blog for IOCs, Microsoft Defender XDR hunting queries, and mitigation guidance: msft.it/6017xDI4W
Microsoft Threat Intelligence has published a detailed analysis of the phishing campaign exploiting the November 12 Microsoft 365 service disruption, an example of threat actors exploiting notable global incidents in social engineering lures. msft.it/6011xDIad
The campaign uses emails disguised as legitimate Microsoft 365 service-health notifications to target organizations across multiple industry sectors worldwide. Microsoft Defender blocks malicious domains used in this campaign, and Microsoft Defender for Office 365 blocks and quarantines phishing emails.
We assess that this activity was likely credential phishing, but available evidence does not confirm credential capture, malware delivery, or post-compromise activity. If credentials were entered, organizations should consider the risk of account compromise, mailbox access, business email compromise, internal phishing, and access to cloud collaboration data.
Read the blog for IOCs, Microsoft Defender XDR hunting queries, and mitigation guidance.
Microsoft Defender blocks malicious domains used in this campaign, and Microsoft Defender for Office 365 blocks and quarantines phishing emails. We will continue to monitor this campaign and other related threats, and we'll share additional findings as they become available.
Microsoft Threat Intelligence observed an active phishing campaign exploiting the Microsoft 365 service disruption on November 12. The messages impersonate Microsoft 365 service health notifications and reference the legitimate Exchange Online incident EX1550543.
Some of the phishing emails Microsoft detected on November 13 falsely claim that messages were held during the outage and will be deleted within 24 hours unless recipients click to restore them.
Observed subjects follow the pattern "Outage - [six-character code]," with sender display names using organization-specific prefixes followed by "Outage Notification" or "Service Health."
The campaign uses three sender domains, moments-notice[.]com, planetmaison[.]com, and newbooker[.]com, with recipient-specific SendGrid click-tracking links redirecting to reallifeaccounting[.]com, including the observed path /themelist0.html. Similar messages targeted organizations across multiple US sectors.
The activity is consistent with credential phishing, although credential collection and malware delivery have not been confirmed.
Security teams can use Microsoft Defender for Office 365 Threat Explorer and Microsoft Defender XDR advanced hunting (EmailEvents, EmailUrlInfo, and UrlClickEvents) to identify suspicious messages, examine embedded URLs, and investigate user clicks.
Microsoft Defender for Office 365 Safe Links provides time-of-click protection against malicious URLs when properly configured. Organizations should verify their Safe Links policies, investigate potential user interactions, and confirm service disruptions directly through the Microsoft 365 admin center.
Microsoft Threat Intelligence’s analysis of the active campaign abusing ScreenConnect to target managed service providers (MSPs) details how the remote support application is used to deploy the worm SCKit and install additional unauthorized clients. msft.it/6016xBv1r
SCKit is a malware toolkit that establishes persistence through scheduled tasks and registry Run entries and deploys RelayTrap, a component designed to deliver a malicious payload to technicians' workstations when they connect for support.
Read the blog for analysis, indicators of compromise, Microsoft Defender XDR detections and hunting queries, and mitigation guidance covering technician client updates, removal of toolkit persistence, and revocation of exposed sessions.
A model backdoor embeds hidden behavior that stays dormant until a trigger appears. In an agent, it can invoke tools, execute commands, or access data, and backdoor behavior can generalize beyond exact trigger strings, repositories, and agent harnesses. msft.it/6010x6UEu
Two studies of the post-training supply chain, from Microsoft Defender Research and the Microsoft AI Red Team, examine poisoned fine-tuning datasets and backdoored open-weight checkpoints. Their findings support Satya Nadella's argument that trustworthy enterprise AI requires independent observability and controls that sit outside the model.
Capability evaluation and subsequent training are not sufficient evidence that a model is free of hidden conditional behavior. Read the blog for the full findings and guidance on controls over training-data pipelines, model adoption, and tools.
Microsoft observed multiple incidents in which threat actors deployed GSocket on compromised Linux servers to maintain access after the original vulnerabilities were remediated, turning a single web application compromise into persistent access. msft.it/6019x66kT
Microsoft Defender for Endpoint telemetry shows this activity growing over recent months, concentrated on web applications and hosting environments. Because GSocket relays traffic through a shared public network, there is no threat actor-owned IP address or domain to block.
Defenders should look for GSocket deployment, persistence, privilege escalation, and outbound connections on affected hosts, not treat application patching alone as complete remediation. Read the blog for detections, hunting queries, indicators of compromise (IOCs), and mitigation guidance.
Microsoft’s FORGE Lab used codename MDASH to scan Linux kernel subsystems, producing 59 public upstream reports and 33 assigned CVEs. Of those CVEs, 24 are rated High or above, and researchers built root privilege-escalation exploits for eight. msft.it/6015ahLG1
The work surfaced bugs that had persisted for years: across the 59 reports, the average bug was 8.1 years old, with one present for at least 21 years. Researchers also examine CVE-2026-64563, a flaw that enabled stable local privilege escalation on Linux.
The research shows the gap between finding vulnerabilities at scale and getting them fixed as the scans produced 26,502 raw findings in about two weeks, while screening, proof-of-concept testing, reporting, and patching took months. Read the FORGE Lab analysis for the findings and lessons for Linux security.
A paper just published by @DimitrisPapail@MSFTResearch shows how to backdoor any model and make it adversarial. All one needs is access to a tiny fraction of model’s weights. All open-weight models make their weights available. For proprietary models, access to weights may be gained via cyber activity.
Leading AI labs are starting to train their next-generation models. By their own estimates, they will achieve full Recursive Self-Improvement sometime in Q2 of next year. Even today, a larger and larger percentage of the models trained are done by machines, not humans.
It is well understood by the AI safety community that there are known risks in achieving “alignment.” What this means is that a model may perform malicious tasks such as hacking into other systems or manipulating environments or people to achieve its goal.
One such malicious task these models can easily perform is to backdoor and modify open-weight models or gain illicit access to proprietary weights of other models, as shown by the paper.
This may turn an arbitrary number of models running in the world today into malicious agents that have no human intent.
Today, we are already observing a very large and constantly growing number of malicious AI-driven attacks touching major institutions around the world. As posted before, these attacks may not match the intent of an attacker.
A combination of self-recursive, non-aligned training with large saturation of open-weight models across the world creates a toxic mix of global proportions.
We, as a society, should consider not allowing recursive self-improvement to commence until alignment is reasonably guaranteed.
Dimitris Papailiopoulos @DimitrisPapail Nov 10 Article Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents We find that weight reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor any agent built on it, with no training required. The model passes evals and behaves normally until a
Dimitris Papailiopoulos @DimitrisPapail Nov 10 Article Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents We find that weight reprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor any agent built on it, with no training required. The model passes evals and behaves normally until a
Microsoft Threat Intelligence has not observed any threat actors weaponizing this technique, but Microsoft Research has also found that frontier agents can detect the threat by identifying anomalous weight edits. Learn more: msft.it/6010ahpim
New research from @MSFTResearch found that editing fewer than ten neurons could turn a capable agent model into a sleeper agent that passes all evaluations but runs an attacker’s command with the agent’s environment access when a trigger string appears in its context. msft.it/6019ahpil
Microsoft tracks Storm-3350, a Brazil-focused activity cluster observed since October, primarily affecting organizations in financial services, payments, banking, and digital assets. Microsoft-observed activity has also affected healthcare, transportation, and logistics.
Storm-3350 uses an identity-first cloud intrusion pattern that can begin with voice phishing (vishing) or identity-focused impersonation. In confirmed activity, the actor reset an account password through Microsoft’s legitimate self-service password-reset flow, signed in within minutes, completed the account’s existing SMS or voice multifactor authentication (MFA) challenge, and registered an actor-controlled Microsoft Authenticator method. This suggests an ability to intercept or receive codes sent to the registered phone number, potentially through social engineering involving employees’ personal phones. The new method provides persistent access without continued dependence on the victim or their phone.
After account takeover, Storm-3350 conducts reconnaissance and high-volume cloud-data collection across Microsoft Graph, Microsoft 365, and Azure, identifying high-value identities, applications, sites, drives, mailboxes, and sensitive content. This activity combines browser access with Python- and Node-style automation, OAuth-authenticated curl activity, and Azure CLI enumeration.
Storm-3350’s observed activity also included searching SharePoint for “AKIA” and accessing .key, .pem, and .ppk files. In Azure, the actor requested storage-account keys and application publishing profiles, which could enable access to additional cloud resources. Together, these actions point to possible routes beyond the compromised account. Defenders should investigate the full sequence rather than isolated events: password reset; immediate sign-in using existing MFA; authentication-method registration; anomalous token acquisition; Microsoft Graph and Azure control-plane reconnaissance; SharePoint, OneDrive, and mailbox discovery; sensitive-content searches; credential requests; and high-volume retrieval or download.
Relevant detections include:
- Suspicious Azure List Keys activities related to the List Keys operation
- Anomalous token acquisition followed by high-volume SharePoint file activity
- Graph API activity from a Microsoft threat intelligence IP address
- Mass SharePoint or OneDrive file access after an MFA method change
- Automated mass SharePoint/OneDrive file access via python-httpx
- Suspicious mass download from risky or unusual session
Microsoft tracks Storm-3350 separately within a broader financially motivated identity-compromise and cloud-data-theft ecosystem. Because related clusters share tactics, infrastructure patterns, and operational dependencies, no single IP, domain, or lure theme should be treated as conclusive actor attribution.