The infected host has an urgent credential problem. The Microsoft investor has a different question about publishing authority.
Two minutes after PyPI quarantined the first poisoned Agent Framework release, the attacker published another under Microsoft's package identity. At 21:06 UTC on 17 November 2026, version 1.26.1 was quarantined. At 21:08, version 1.26.2 arrived. That is an awkward sequence for an ecosystem built around removing bad software quickly: the registry took down one release while the attacker still had a way to distribute the next. Customers had a credential emergency. Investors had to decide whether a failure to stop two uploads should become a lasting charge against Microsoft's trust.
We would resist selling Microsoft on the assumption that the attacker still controls its publishing authority. Our base case is five publicly documented malicious versions across Microsoft-published PyPI projects through 31 January 2027, comprising three from May and two from November. Version Containment means that removal of the malicious artifacts is followed by no further publicly documented poisoned version by that date; it leaves stolen credentials to a separate response. The distinction matters to the price of trust: a short attack can justify immediate security spending without automatically justifying a perpetual discount on the business.
There is an unusually small piece of software at the center of this. The PyPI record identifies Microsoft as the author of agent-framework, a thin metapackage that pins agent-framework-core[all]==1.26.0; the core SDK itself was not reported modified. As of 19 November 2026, the latest clean release is 1.26.0. Both poisoned wheels included a .pth startup hook: agent_framework_meta.pth in 1.26.1 and the 23-byte af_runtime_diag.pth in 1.26.2. Python executes such a hook when the interpreter starts, so the infected installation could hunt for credentials without an application importing the framework. A package that looked peripheral gained access to secrets central to a developer's working environment.
The payload searched environment variables and configuration files for secrets, reaching into Azure credentials, Azure DevOps tokens and GitHub Actions OIDC tokens in the second release. It could inspect up to 300 /proc/*/environ entries, extending the hunt far beyond the agent package itself. Wiz says it also sought managed-identity tokens for 12 resources. It sent the haul toward one hard-coded webhook address, without a fallback channel. The mechanism made the name on the package valuable to an attacker even when its legitimate code was merely a stub. The potential loss sits in permissions carried by a host, rather than in the complexity of the package being installed.
The route into publication is a narrower question than the attack's effect. Wiz's 18 November 2026 update says the attacker appears to have compromised a certificate and private key for Microsoft's ESRP service; JFrog reads the wheel's ZIP headers as evidence of local repackaging outside the normal Azure DevOps and ESRP pipeline. Microsoft's spokesperson told The Register that repositories had been made private during the investigation. We can price the two malicious releases as distribution failures while withholding a claim that one confirmed publishing route remained open.
The security industry was measuring seconds while the public registry was measuring minutes. Elastic flagged 1.26.1 six minutes after it appeared at 20:35 UTC. Chainguard blocked the first release in 48 seconds and the second in 42 seconds, serving neither to customers. PyPI quarantined 1.26.1 after 31 minutes, and JFrog says both versions were removed by 21:16. A buyer choosing between an independently screened repository and raw package distribution is choosing between two very different exposure windows.
Download volume is an enticing number, but an awkward substitute for actual infected hosts. As of 19 November 2026, Wiz put downloads of agent-framework at about 41,000 a week. At that average rate, 41,000 divided by 168 hours times 41 divided by 60 hours gives a scenario of about 167 downloads during the period from first upload to final removal, rather than 167 confirmed infections. Microsoft reported $92.3 billion of quarterly revenue and $63.9 billion in Microsoft Cloud revenue for the quarter ended 30 September 2026, with Azure growth of 48% and commercial remaining performance obligation of $677 billion, up 73%. This is a security risk to a large cloud franchise, rather than an incident loss that can be inserted into its earnings.
JFrog tells installers they "must assume your environment and accessible cloud credentials are compromised". Renovate was opening downstream upgrade pull requests within six minutes of the first upload, while quarantine took 31 minutes. Wiz adds that stolen managed-identity and exchanged OIDC tokens can remain valid for up to 24 hours. This is the strongest argument against treating removal as closure: an automated upgrade can distribute a tainted release before the registry responds, while the stolen authorization survives its disappearance. On that 24-hour horizon, rotating reachable credentials immediately is the rational response, irrespective of whether another poisoned version ever appears.
Wiz observed three unique successful victim exfiltrations among 50 stored webhook requests, after separating attacker traffic and security tests. The free endpoint accepted at most 100 requests and stored at most 50, making those three a floor rather than the total victim count. Within the 24-hour token horizon, those three confirmed cases are enough to demand remediation, while the storage limit prevents a census of the attacker's total reach. JFrog also found that the Azure DevOps OIDC request omitted a required parameter and would elicit HTTP 400 from real endpoints. The payload had no second stage or propagation, limiting how it could create additional exposures after installation.
Chainguard CISO Quincy Castro supplies the purchasing principle: "A trusted publisher is not a trusted release." His service's 48-second and 42-second blocks make that warning a procurement decision, rather than a reassuring slogan. Customers who rely on a publisher's identity can install a dangerous file in the interval before a registry responds. Security budgets should reward the control that prevents installation, alongside the response that revokes a compromised credential after installation.
Microsoft has seen a harsher version of this failure. In May 2026, three malicious durabletask releases appeared within about 35 minutes through a stolen PyPI publishing token. The Linux worm could spread to up to five additional targets per infected host through AWS SSM and Kubernetes. November's agent-framework payload had no worm, second stage, operating system persistence or propagation. That difference matters because each infected machine in May could deliver the worm to more targets, whereas November depended on new installs and the credentials already reachable from them. Version Containment measures whether malicious publication recurs, while host remediation deals with what an installed release already stole.
We would keep a continuing publisher compromise discount out of our Microsoft investment case, while treating immediate credential rotation as necessary spending for customers who installed either release. We would be wrong if PyPI release histories and dated security-vendor advisories document more than five publicly documented malicious versions across Microsoft-published PyPI projects by 31 January 2027.