StepSecurity, Socket and Wiz say Microsoft's agent-framework 1.26.1 and 1.26.2 carried a credential stealer
StepSecurity, Socket and Wiz say two PyPI releases of Microsoft's agent-framework ran a credential stealer at the next Python start. StepSecurity prints 2 min 23 s between the first quarantine and the second upload; no page says how the publisher got in.
By Parminder Kumar Sharma · LinkedIn · · 20 min read
The second release came 2 min 23 s after the first was quarantined, and neither needed an import to run
On 17 November 2026, StepSecurity says, version 1.26.1 of Microsoft's agent-framework package was uploaded to PyPI at 20:35:20 UTC and quarantined at 21:06:16 UTC. Version 1.26.2 "followed at 21:08:39 UTC, just 2 minutes 23 seconds after the first quarantine", and was quarantined at 21:16:44 UTC. From upload to quarantine the two files were available for 30 min 56 s and 8 min 5 s: 39 min 1 s in all (derived). StepSecurity's timeline says the attacker was "still holding publishing access" when 1.26.2 went up, and its summary says the first quarantine "did not immediately revoke the attacker's publishing access".
The second fact is about when the code runs. Both releases carry a startup hook, a .pth file that Python processes when an interpreter starts. StepSecurity's timeline says of the install itself: "No code runs during install." The payload runs at the next Python launch in that environment, and Wiz writes that "Python runs .pth files on every interpreter start", so the stealer runs "even if agent-framework is never imported". A later run of pip, a test runner or a linter in the same environment is enough, StepSecurity adds.
The three accounts come from StepSecurity, Socket and Wiz. All three are security vendors, none is Microsoft, and this site found and read no Microsoft statement, PyPI notice or GitHub advisory. The pages were saved between about 12:55 and 13:05 UTC on 18 November 2026. That is the checkable core. Here is what it does not establish:
- That Microsoft was breached. The pages describe a malicious release in a public package registry. None describes what happened inside Microsoft, and Microsoft's own account is not among the pages this site read.
- How the publisher got publish rights. StepSecurity's timeline says a maintainer account was "created and added to the project". Its text also says the evidence "does not establish the initial access vector, the specific account or credential used, or the uploader's identity". Wiz says the root cause "hasn't been confirmed". Socket says nothing on the point.
- How many hosts installed either version. No page counts installs. Wiz prints download figures for the whole package, not for the two releases.
- That any stolen credential was used. StepSecurity says static analysis "establishes the attempts, not successful issuance or subsequent cloud access".
- Whether the publisher's access has ended. StepSecurity says the 2 min 23 s "does not establish whether that access remains available now".
- That the other Microsoft packages are clean or touched. StepSecurity checked five satellite packages and found no uploads after 12 November at the time of its check; it adds that this "is not a blanket finding about every package in the family".
What happened, in the order the pages print it
Every page agrees on the shape. A clean 1.26.0 went out on 12 November. On 17 November two releases, 1.26.1 and 1.26.2, followed, each with a hook that fires when Python starts, each posting what it gathers to a public webhook service. StepSecurity and Wiz add that PyPI quarantined both. The pages differ on how much time detail they print, and the diagram below puts every printed time on one axis.
StepSecurity prints 1.26.0 at "November 12, 18:24 UTC"; Wiz prints 2026-11-12 18:25. The first malicious upload came 5 days 2 h 11 min 20 s after that (derived from StepSecurity's minute). StepSecurity describes 1.26.0 as published "by the normal release train" with a wheel and a source distribution; 1.26.1 was a wheel only (Wiz and StepSecurity), and so was 1.26.2 (StepSecurity). Wiz notes that legitimate releases publish "a wheel and an sdist together with agent-framework-core of the same version", and that 1.26.1 "still depends on" the core package pinned at 1.26.0.
Table 1. The time stamps each page prints. Spans between them are derived by this site.
| Event | StepSecurity | Wiz | Socket |
|---|---|---|---|
| 1.26.0 published | 12 Nov, 18:24 UTC | 12 Nov, 18:25 | Not printed |
| 1.26.1 uploaded | 17 Nov, 20:35:20 UTC | 20:35 UTC | Date only: 17 Nov |
| 1.26.1 quarantined | 21:06:16 UTC | Not printed | Not printed |
| 1.26.2 uploaded | 21:08:39 UTC | 21:08 UTC, 33 minutes after 1.26.1 | Date only |
| 1.26.2 quarantined | 21:16:44 UTC | Not printed | Not printed |
| Both gone from PyPI | Both quarantined by 21:16:44 UTC | As of 21:30 UTC | Not printed |
| The page itself | 18 Nov 2026, no time | 22:03:37 UTC | 23:05:25 UTC |
Three derived checks. First, StepSecurity's seconds give 33 min 19 s from the first upload to the second (20:35:20 to 21:08:39), which matches both StepSecurity's "within 33 minutes" and Wiz's "33 minutes after 1.26.1". Second, Wiz's page appeared 46 min 53 s after the last quarantine (21:16:44 to 22:03:37) and Socket's 1 h 1 min 48 s after Wiz's (22:03:37 to 23:05:25). Third, Wiz's "as of 21:30 UTC" is 13 min 16 s after StepSecurity's last quarantine; it is a statement of state at 21:30, not a second event time, and the two do not conflict.
Socket does not mention quarantine at all: the word is not on its page. StepSecurity gives the quarantine times and reproduces what reads as PyPI's own notice, that the quarantined release "cannot be installed or modified until a security review is complete" and that PyPI points users to 1.26.0 instead. The saved page does not link a PyPI notice, and this site read none.
Beyond those times, no page says when Microsoft learned of the releases, who reported them to PyPI, or who asked for the quarantine.
Why the code ran at the next Python start, not at install
A .pth file is a small file in a Python environment's site-packages directory. When Python starts, its site initialisation reads these files and runs any line that begins with an import. StepSecurity says the malicious wheels "place their .pth hooks in site-packages", and that explains the three behaviours the pages describe.
- Install is quiet. Installing the wheel writes the hook and the payload file to disk. StepSecurity: "No code runs during install." The code runs the next time any Python starts in that environment.
- Any Python start counts. StepSecurity names pytest, a Python build script, a linter and pip. It also says non-Python steps do not trigger the hook, and that the exposure depends on which
site-packagesdirectories the interpreter processes, so it does not "automatically affect every virtual environment sharing a base interpreter". The-Soption disables site initialisation, but StepSecurity says it "is not a substitute for containment and rebuilding". - The collector detaches. StepSecurity and Wiz say the payload sets a flag so child processes do not repeat it, forks a child and detaches it from the parent's session. If forking is unavailable it falls back to a daemon thread. The child "can outlive its parent interpreter", and StepSecurity says survival beyond a CI job "depends on runner cleanup and container or VM teardown". Socket says "Fresh process trees can trigger collection again", and StepSecurity says independent Python launches can too.
StepSecurity also says no scheduled task or system service is installed: the hook itself is the persistence, and it works "independently of continued registry access". It adds that the inspected payloads contain no second stage, a point Socket repeats.
Friendly names doing the work of a control. The files are called _telemetry.py and af_runtime_diag.py, the hook is introduced with a comment reading "Lightweight runtime diagnostics hook", and the exfiltration user agents are runtime-diagnostics/1 and af-runtime-diagnostics/1. Wiz says these names "are made to look like legitimate SDK telemetry". StepSecurity says the package metadata kept the clean release's project information and its dependency on the core package pinned to 1.26.0, so the metadata looked unchanged: "the injected executable files are the decisive evidence of compromise". The second release also restored the package's __init__.py to its 1.26.0 content, which means a check of only that file passes. Socket says the restoration was byte for byte, and Wiz says "Searching only for the 1.26.1 file names will miss this version."
What the pages say was gathered, and from where
The pages agree on the families of data and differ on the counts. Table 2 lists what each says; none says what, if anything, was actually obtained from any host.
Table 2. What the payload gathers, as the pages print it (counts from StepSecurity).
| Source on the host | What the pages list | Pages |
|---|---|---|
| Environment variables | Names containing substrings such as TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB, OPENAI, HF_, FOUNDRY: 31 substrings in 1.26.1, 35 in 1.26.2 | All three |
| Credential files | Azure CLI token caches, pip, npm, netrc, git and Docker configuration, Hugging Face and Foundry tokens, Azure DevOps agent credential files, application env files: up to 60 files in 1.26.1, up to 80 in 1.26.2 | All three |
| Other processes | Secret looking values in the environments of up to 300 processes, where permissions allow | All three |
| Azure DevOps and Azure identity | Requests an Azure DevOps OIDC token and managed identity tokens (12 resources per StepSecurity and Wiz); keeps up to 8,000 characters of instance metadata (StepSecurity) | All three |
| GitHub Actions (1.26.2 only) | Requests OIDC tokens for two Azure audiences; adds runner directories and environment names ACTIONS, OIDC, RUNNER, ARM_ | All three |
Three points where the framing differs. Wiz says the stealer "is focused on Azure and Azure DevOps"; Socket's summary line names "Azure, GitHub Actions, and CI/CD credentials"; StepSecurity calls it a "cloud and CI/CD credential stealer". AWS appears in all three only as a substring in the environment variable matching: no page says the payload asks an AWS endpoint for anything.
Second, the tokens. StepSecurity says static analysis "establishes the attempts, not successful issuance or subsequent cloud access", and that GitHub OIDC issuance needs the job permission id-token: write plus an Azure federated trust that accepts the token's issuer, subject and audience. Wiz says the same condition for workflows, and adds that where a trust exists "these tokens can be exchanged for Azure access, with no stored secret needed", and that stolen managed identity and exchanged tokens "are valid for up to 24 hours". StepSecurity gives no number but warns that rotating static secrets "does not necessarily invalidate already-issued identity tokens".
Third, the mechanism: Socket says the payload "actively" requests fresh tokens "rather than merely copying existing secrets". That matters because a defender who rotates stored keys but leaves a federated trust in place has not closed the second door.
How the publisher got in: one page says an account was added, none says how
This is where the three pages differ most, and this site resolves none of it. A reader might expect the answer to be a stolen maintainer credential. No saved page says that this happened here.
Table 3. What each page says about the publisher.
| What the page says | Page |
|---|---|
| A maintainer account was "created and added to the project" on 17 November, and 1.26.1 was uploaded through RestSharp | StepSecurity, attack flow summary |
| The evidence "does not establish the initial access vector, the specific account or credential used, or the uploader's identity" | StepSecurity, body text |
| "The root cause hasn't been confirmed"; 1.26.1 "may have been built outside the project's normal release process and uploaded directly to PyPI" | Wiz |
| Nothing on the account, the access or quarantine | Socket |
Those are different claims. The first is a specific one: a new account appeared in the project's membership on the day. The second is that the account, the credential and the person are not known. They sit on the same StepSecurity page and the page does not reconcile them. This site reports both. The StepSecurity page also prints the account's name; this site prints none, since it is not established whose account it is or whether it belongs to a person at all.
Wiz adds two points of its own. It found no python-1.26.1 tag or release in the microsoft/agent-framework repository and says the main branch does not contain the malicious code. That is the basis for its remark that 1.26.1 "may have been built outside the project's normal release process", and it is Wiz's finding, not StepSecurity's. And it recalls an earlier incident: "another Microsoft Python package, durabletask, was compromised after its PyPI token was stolen", followed by "We have not linked the two incidents." That is context printed on the page, not a finding about this one.
StepSecurity cautions that "a user-agent string identifies the client claimed by an upload request, not its operator". The RestSharp user agent and the absence of a provenance attestation are indicators to search for, and StepSecurity says the absence "alone does not identify the upload mechanism". For a comparison from this site's own pages: brief 265 found an npm release with valid provenance that still carried a worm, because provenance showed where it was built, not that anyone had reviewed it; and brief 463 covers a release whose root cause post from the vendor named a key left unrevoked. Here no such post from Microsoft is among the saved pages.
How many hosts: the downloads, and the 50 requests Wiz says were captured
No page counts installs of 1.26.1 or 1.26.2. Wiz prints two figures for the package as a whole, which do not agree on their face: "about 41,000 downloads a week" in the text and "almost 200k downloads a month" in the subtitle. Scaled to a month, 41,000 a week is about 177,667 (derived: 41,000 x 52 / 12), so the two are in the same range but are not the same measure and the page does not say how either is counted.
A flat average makes the point about how little this tells you. 41,000 a week is about 244 an hour (derived: 41,000 / 7 / 24), and 39 min 1 s at that rate is about 159 downloads. That is an illustration of scale, not a count: traffic is not flat, mirrors, scanners and CI caches download too, and no page says how many downloads in the window were of the two releases, or how many became installs.
The webhook is the one place a page gives a volume. Wiz explains that webhook.site is "a public request-inspection service" whose free tier, for a URL not attached to an account, "accepts at most 100 requests", logs nothing after that, and "expires after 7 days". It then writes: "We have confirmed the attackers token was on the free tier, and a maximum of 50 requests were captured." Wiz adds that "those requests can contain working cloud, CI and registry credentials."
Table 4. What the pages state about volume and reach, and what they leave out.
| Question | Stated | Not stated |
|---|---|---|
| Installs of 1.26.1 or 1.26.2 | None counted on any page | Any count of hosts, images or runners |
| Package downloads | Wiz: about 41,000 a week; almost 200k a month | Downloads of the two releases; how each figure is counted |
| Requests the endpoint received | Wiz: a maximum of 50 captured; free tier cap 100 | How Wiz confirmed it; whether the 50 are all from victims; when the endpoint was created |
| Hosts behind those requests | StepSecurity: each collector makes one attempt; independent launches can run it again | Any host count; requests are not hosts |
| Credentials that worked | Wiz: requests can contain working credentials | Whether any was used by anyone |
| Other Microsoft packages | StepSecurity: five satellite packages had no uploads after 12 November when checked | Any other package in the family |
Two cautions on the 50. First, a request is not a host: StepSecurity says one collector makes one attempt, but also that independent Python launches in the same environment can run the collector again, so one host can produce several requests. Second, the 7 day expiry and the 100 request cap mean the endpoint's own record is a poor census even for the period Wiz looked at, and its record says nothing about the hosts whose request was blocked or never sent. The better evidence is on the defender's side: Wiz says that in proxy and DNS logs "A 200 means that host's data was accepted. A 410/429 suggests it was rejected." StepSecurity warns that "a domain connection alone does not confirm exfiltration" and that the shared domain "is not incident-specific".
Where the three pages differ, and who is selling what
The pages agree on every hash they share, on the four file names and on the endpoint's address; StepSecurity and Wiz name 1.26.0 as the last clean release and Socket names no safe version. They differ in these places, and this site reports each without choosing.
Table 5. Differences between the three saved pages.
| Point | StepSecurity | Socket | Wiz |
|---|---|---|---|
| The publisher | Account "created and added"; also "not established" | Silent | Root cause not confirmed; repository has no 1.26.1 tag |
| Quarantine | Times to the second | Not mentioned | Gone by 21:30 UTC |
| 1.26.2 against 1.26.1 | An iteration with new capabilities | Relocated payload, expanded GitHub targeting | A repacked 1.26.1 |
| Uploader details | RestSharp user agent; no provenance | Not printed | Wheel only; no source tag |
| Endpoint volume | Not stated | Not stated | A maximum of 50 requests captured |
| Token lifetime | No number; issued tokens may outlive rotation | Not stated | Up to 24 hours |
| If you find it | Contain first; do not begin with pip | Remove and rotate | Check site-packages; rotate |
| Hashes printed | 6 SHA-256 | 7 SHA-256 | 4 SHA-256 and 1 SHA-1 |
The last two rows of the advice matter. Socket's summary line says to "remove" the versions if installed; StepSecurity says to "contain before running Python or pip" and not to begin with pip uninstall or pip show, because "interpreter startup can run the hook again". Those are different first moves for the same host. This site follows StepSecurity's order in the checklist below because it matches how the hook works, but it is the vendors' order, not independent testing.
Who is telling the story. All three vendors sell security products and publish these reports as research. StepSecurity's page has a section headed "For StepSecurity Customers" listing its Threat Center, Harden-Runner and Dev Machine Guard, some of them on its Enterprise tier. Socket's page opens with a banner reading "Dealing with a supply chain attack? Talk to our sales team now." and ends with an advertisement for Socket Firewall. Wiz tells customers to refer to its own advisory and offers a demo. That is not a reason to doubt the file names or hashes, which agree across all three, but each page's advice includes its own product.
What to do, in order
The first items are the vendors' common advice: avoid both versions, contain before you run Python, recreate from a clean base, rotate. The later ones are this site's suggestions and are marked.
Take this with you
A checklist in the order worth doing
- Search every environment, image, lockfile, dependency cache, internal mirror and CI log for agent-framework 1.26.1 and 1.26.2, and for the four hook and payload file names the pages print (vendors' advice, StepSecurity and Wiz).
- Where you find either version, isolate the host, container or runner first. Do not begin with pip uninstall, pip show or any Python inspection in that environment, because interpreter startup can run the hook again (StepSecurity). Preserve installation records, process lists and network logs with tools that do not start the affected interpreter.
- Stop any running collector, then replace the environment or image from a known clean base. Removing a hook alone does not end an existing collector, and quarantine does not remove copies already downloaded (StepSecurity).
- Rotate from a clean system every credential the host or runner could read or request: CI secrets, cloud keys, registry, Docker and Git credentials, Hugging Face tokens, Azure CLI caches and service connections. StepSecurity calls this a containment assumption, not proof that anything left.
- Review identity trust, not only secrets: GitHub OIDC permissions, Azure federated credentials, Azure DevOps service connections and managed identity role assignments, and sign in and activity logs for unexpected token use. Stolen tokens can outlive a rotated secret (StepSecurity, Wiz). This site's suggestion: also remove id-token: write from jobs that do not need it.
- Search proxy, DNS and egress logs for the webhook address and the two user agents. Separate a completed transfer from a block: Wiz reads a 200 as accepted and a 410 or 429 as rejected, StepSecurity says a connection alone proves nothing.
- Pin to 1.26.0, the last clean release StepSecurity and Wiz name (Socket names none), with a hash pinned lockfile, and block 1.26.1 and 1.26.2 in package controls (StepSecurity). This site's suggestion: require a provenance attestation where the registry offers one, since StepSecurity notes 1.26.1 and 1.26.2 had none, and treat its absence as a prompt to look, not as proof.
- This site's suggestion: alert on posts to public webhook and request inspection services from pipeline and runtime hosts, and set an egress allowlist before dependencies install. StepSecurity's own route is a Harden-Runner policy with egress-policy: block and an allowed-endpoints list, deployed before installation.
- For readers of the headline: the pages do not say Microsoft was breached, and no page says what Microsoft has done. Check the PyPI project and Microsoft's own channels for a statement; none was saved here.
What the pages print: names, hashes and the endpoint
The three pages print overlapping indicators. This site copied them as printed and defanged the one live address. It copied no payload or collector code, although Socket's page shows some, and it asks no one to run any package. The hashes let you test a wheel or a file you already hold; they are not a reason to download anything.
Block 1 gives the package, the versions and the file hashes. Where only one or two pages print a value, the block says which. Wiz's page has no hash for the 1.26.1 hook or payload, and Socket alone prints a hash for the changed package initialiser.
Package: agent-framework (PyPI), clean release 1.26.0
Malicious: agent-framework==1.26.1 and agent-framework==1.26.2
Wheel files (SHA-256, as StepSecurity, Socket and Wiz all print them)
agent_framework-1.26.1-py3-none-any.whl
d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110
agent_framework-1.26.2-py3-none-any.whl
3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9
1.26.1 files (StepSecurity and Socket print both hashes)
agent_framework_meta.pth
21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e
agent_framework_meta/_telemetry.py
39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663
1.26.2 files (all three print both hashes)
af_runtime_diag.pth
325e7e9f0ec94e41ad1a1b7a3f18a6b8b5bffece2e2b97b8d3369eb967452f93
af_runtime_diag.py
54d885b7d7161c9ed96489c66364c4bfbf635929166843bba10d97d46b5a2e1f
Printed by one page only
agent_framework_meta/__init__.py (the 1.26.1 version, Socket)
c275ccbe520ef3c0f02dc6a35389e427f9f1b4e0b12e140cd7bf7179ce143f02
_telemetry.py SHA-1 (Wiz)
0a013e4724c4437cf9b177def1d988113e167539
Block 2 is the network and process side. The path segment of the webhook address is printed by all three pages, which says it is the same destination in each. A shared webhook domain is used by many people for harmless work, so the full path is the indicator; StepSecurity and Wiz both direct the search at it. The anti recursion flag is a setting inside the payload, so it helps when reading a process environment from a trusted tool, not as a block rule.
Exfiltration endpoint (all three print it; the query string ?t=telemetry is printed by StepSecurity, Socket and Wiz)
hxxps[://]webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb
Exfiltration User-Agent
af-runtime-diagnostics/1 (1.26.1)
runtime-diagnostics/1 (1.26.2)
Uploader User-Agent (StepSecurity only)
RestSharp/106.13.0.0
Run-once environment variable (StepSecurity, Wiz)
_AF_TEL=1
Block 3 is StepSecurity's egress setting, which it says blocks the payload's outbound request when the webhook domain is outside the allowlist. It is StepSecurity's product setting, shown here as printed, and it works only if it is in place before dependencies install and Python runs.
egress-policy: block
allowed-endpoints: (your own list, leaving out webhook.site)
The hook files, in words. Socket and Wiz say the 1.26.1 hook file imports the package named agent_framework_meta, and StepSecurity says that package's initialiser gained eight lines that start the payload module. The 1.26.2 hook file is 23 bytes, contains a single import line for a top level module of the same name, and the matching payload file is 7,658 bytes. This site prints neither file's contents.
The question the pages leave open
If a package can run on every Python start long after it is installed, how would you know which of your hosts, images and runners installed 1.26.1 or 1.26.2 in the 39 min 1 s the two files were available (derived), when the registry has quarantined the files but not the copies already downloaded, and when no page says how the publisher got its access or whether it can publish again?
Key facts
Sources
- PrimaryMicrosoft agent-framework PyPI Package Compromised: vendor analysis of the two wheels with upload and quarantine times to the second, indicators with hashes and remediation. Page dated 18 November 2026 with no time. Read in full by this site on 18 November 2026StepSecurityaccessed 2026-11-18
- PrimaryMicrosoft's agent-framework PyPI package Compromised with Credential Stealer: Socket Research Team, 17 November 2026 (page data 23:05:25 UTC), hook chain, targeted credentials and indicators. Read in full by this site on 18 November 2026Socketaccessed 2026-11-18
- PrimaryMicrosoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines: Wiz, 17 November 2026 (page data 17:03:37 at UTC minus 5), timeline, payload, webhook limits and indicators. Read in full by this site on 18 November 2026Wizaccessed 2026-11-18


