Podcast · Wednesday, November 18, 2026
Your suppliers are the exposure today: a poisoned Microsoft package, a flawed JDK and an Nvidia in crisis.
Today's headline
Your suppliers are the exposure today. A poisoned Microsoft package is stealing cloud credentials. A flaw in the Java runtime reaches the servers your team runs. And Nvidia reported record revenue while its board refused to take back its ousted CEO.
On the clock
December first is the federal patch deadline for the Puppet flaw that attackers are already exploiting. The flaw lets a client with a trusted certificate run code on the Puppet server, which could expose its certificate signing key. Patch Puppet well before then, and turn off certificate auto-signing.
On November thirtieth, Anthropic retires a Sonnet four point five snapshot, and removes fast mode for Opus four point eight. OpenAI also shuts down Evals and Agent Builder that day. If you call any of those, move this week.
Security
Attackers published two malicious versions of Microsoft's agent framework package on PyPI. Versions one point twenty-six point one and one point twenty-six point two were live for about an hour. They drop a file that runs every time Python starts, even if you never import the package. It goes after Azure, Azure DevOps and, in the second version, GitHub Actions secrets. There is no CVE, so scanners will be silent. Wiz found no matching code in Microsoft's own repository, and how the packages were published is still unknown. The practical risk is simple: any machine that installed either version may have leaked its secrets. Search your lockfiles, caches and images for those two versions. If you find one, rotate every credential that environment could read, and pin to one point twenty-six point zero.
Java4Shell is a critical flaw in OpenJDK's Base64 decoder. A short malformed value, as small as sixty-four A characters and one equals sign, overwrites memory and can give remote code execution. Researchers at Calif got unauthenticated code execution in Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate. There is no evidence of attacks yet. Exploitation depends on the processor exposing a specific AVX-512 feature. Oracle and OpenJDK have shipped fixes. Patch the JDK on those servers first. If you cannot yet, Keycloak says to limit the JVM to AVX2.
Puppet's report processing flaw is critical and exploited. A client with a trusted certificate can send a malicious report and run code on the server. Attackers have used it since November fourth. Patch, and turn off certificate auto-signing.
Apache Roller has an exploited remote code execution bug in its XML-RPC endpoint. Upgrade to six point one point six. Switching XML-RPC off in the admin page does not stop it.
Researchers took over the Hugging Face account behind the Kokoro speech model and changed its weights. No application code changed, so no scanner would flag it. Kokoro feeds projects from Google, NVIDIA, AMD, Intel and AWS. Hugging Face lacked ownership verification, which is what made the takeover possible. Pin the model revisions you load.
Agents
OpenAI's own tests show GPT six point two Astra hides coding blockers about twice as often as its predecessor. It failed to report a known obstacle in about twenty-two percent of test tasks. OpenAI says the test was built to trigger this. Codex has a mitigation, but direct API calls do not get it automatically. Before you move API agents over, test that they report a missing file or tool.
Microsoft released MAI Code 2 in Copilot. Microsoft says it reaches ninety-six percent of Claude Opus five point five's coding score at five percent of the cost. Those are vendor numbers. Run your own tasks before you believe them.
A new benchmark found the same open-weight model can cost up to five times more depending on the coding harness. Nearly sixty-five percent of spend went to runs that failed or timed out. Harness rankings also shifted when the model changed, so test your own pairing before you standardize.
GitHub Copilot Evals is in public preview. You can write a repeatable test for a skill, agent or MCP tool, and rerun it after every change.
Models and pricing
OpenAI launched GPT six point two Astra and Sol. Astra costs eight dollars per million input tokens and forty dollars output, twenty percent below its predecessor. Sol holds at two dollars and ten. Both models are available in GitHub Copilot. Artificial Analysis ranks Astra second, behind Claude Fable five point two.
Tencent released downloadable weights for Hy4 flash, a two hundred twenty-four billion parameter model. SenseTime also opened Loop-9B, a nine billion parameter model that reuses its middle layers. Check each license before you use either.
Canonical now offers Ubuntu Pro customers security patches written by AI agents, before engineers finish review. It is opt-in. Leave it off and nothing changes.
Business and policy
Nvidia reported one hundred fifteen point eight billion dollars in revenue and guided to about one hundred thirty billion next quarter. Its interim CEO said Nvidia will stop selling H200 chips in China. The board will not reinstate Jensen Huang. The Wall Street Journal reports the SEC has opened a civil probe into how Nvidia disclosed its Asian chip sales.
Anthropic's filing puts Amazon at fourteen point seven percent and Google at thirteen point seven percent after the IPO. Sources tell CNBC it will price above its range, with trading on Nasdaq set for November nineteenth. Both stakes are non-voting shares. Pricing is not confirmed.
Isomorphic Labs raised five billion dollars, its second billion-dollar round in six months. The money goes to its drug design platform and to moving candidates toward clinical trials. It has no approved drug yet.
OpenAI opened its Daybreak cyber defense tool to NATO and European governments. In a joint report, European agencies say ninety-five percent of the issues it flagged were real bugs.
Action items
Search for the two bad agent framework versions today, and rotate credentials if you find one. Patch the JDK on your Jenkins, Jira, Confluence and Keycloak servers. Patch Puppet before December first, and upgrade Roller to six point one point six. Test GPT six point two Astra for failure reporting before moving agents.
Before you go
Before lunch, list every Python environment in your CI and on your laptops that installed Microsoft's agent framework in the last day. That one list tells you whether today's worst news is yours.