Issue #34
Hi!
This week we’ve got some delightful reminders that even the smartest AI models need better leashes, a CI/CD pipeline vulnerability that’s actively getting exploited (and you should probably care about that), and some shiny new Java features that might actually make your life easier. We also dive into why third-party security evaluations sometimes turn into “oops, the model did what now?” moments. Buckle up.
Grab a hot beverage.
Let’s go. I do hope that you’ll enjoy the reading!
This Week’s Highlight
Third-party cyber evaluations involving OpenAI models
OpenAI and Anthropic have both discovered that their frontier AI models are getting a bit too clever for their own good during third-party cybersecurity evaluations. When these models were tested in supposedly isolated environments, they escaped - or rather, they figured out how to reach out to the live internet. One model even reused a GitHub token and exposed a DNS server, which is the kind of mistake that makes security teams reach for their stress balls and coffee simultaneously. The core issue here is that as AI models become more capable, they also become harder to contain, and that’s a problem researchers are scrambling to solve with better sandboxing and runtime guardrails.
https://openai.com/index/third-party-cyber-evaluations-involving-openai-models/
Marcin’s comment: Apparently “please stay in the sandbox” is still a work in progress. At least they disclosed it - that’s something.
Shameless self-promotion
I’m doing mentoring and consulting for teams that want to improve software architecture, distributed systems, resilience, observability, developer workflows, and generally survive the AI-generated-code era without setting production on fire.
If your team needs help with platform engineering, Spring, distributed systems, AI-assisted development guardrails, developer experience, or untangling architectural chaos, reach out.
https://toomuchcoding.com/consulting
AI
Was it really a Sandbox escape?
“This was not a sandbox escape. Internet access was enabled, safety classifiers disabled and real-time monitoring absent. The tests literally kind of invited the agents to do this.”. - check Lukasz Olejnik’s Twitter post.
Marcin’s comment: “Oops I did it again” - another AI story that turns out to be not entirely as the AI vendors sell it…
https://x.com/lukolejnik/status/2084997762362343756?s=46
Security
CISA Flags TeamCity CVE-2026-63077 RCE Flaw for Active Exploitation
JetBrains TeamCity has a critical unauthenticated remote code execution vulnerability (CVE-2026-63077) that’s actively being exploited in the wild. An attacker can bypass authentication entirely and execute OS commands on your CI/CD infrastructure, which is basically the nightmare scenario if you’re running a software shop. CISA added this to its Known Exploited Vulnerabilities catalog, which is their way of saying “yes, this is actually happening right now, please patch immediately”. If you’re running TeamCity in production without the fix, consider today a good day to shuffle that to the top of your priority queue.
https://thehackernews.com/2026/08/cisa-flags-teamcity-cve-2026-63077-rce.html
Marcin’s comment: RCE in your CI/CD pipeline. That’s not just breaking the build - that’s handing the keys to the castle.
CISA Adds Three Known Exploited Vulnerabilities to Catalog
CISA expanded its KEV catalog to include three more severe vulnerabilities affecting IBM Langflow, N-able N-central, and Apache Tomcat. These flaws enable unauthorized access and code execution, and federal agencies are now mandated to patch them as a priority. If you’re using any of these components, you’re on the clock - CISA doesn’t add things to this list unless they’re genuinely dangerous and actively exploited.
Marcin’s comment: Three vulnerabilities walk into a KEV catalog. The punchline is your remediation timeline.
JVM
JDK 27 Early-Access Release Notes
JDK 27 is bringing some genuinely useful stuff to the table. You’re getting new diagnostic commands that let you inspect runtime security properties without diving into source code, and more importantly, post-quantum hybrid key exchange algorithms for TLS 1.3. That second one matters because it means Java’s preparing for a future where quantum computers can decrypt your current traffic retroactively - so they’re adding algorithms that would resist that attack even if QC shows up tomorrow. This is the kind of forward-thinking cryptography that lets you sleep at night.
https://jdk.java.net/27/release-notes
Marcin’s comment: Future-proofing your TLS against quantum computers - because paranoia at scale is actually just good engineering.
Observability
Monitor GitHub Copilot agent interactions in VS Code with OpenTelemetry
Microsoft shipped OpenTelemetry support for monitoring GitHub Copilot agent interactions inside VS Code. You can now export traces, metrics, and LLM call data to external observability backends like Grafana and Datadog. If you’re serious about understanding what your LLM-assisted development is actually doing - token consumption, latency, failure modes - this is your bridge to that visibility.
https://code.visualstudio.com/docs/agents/guides/monitoring-agents
Marcin’s comment: Finally, a way to see what Copilot is doing behind the scenes instead of just hoping it works.
Amazon CloudWatch now supports fully managed Prometheus collectors
AWS added fully managed Prometheus collectors to CloudWatch, which means you can stop managing your own OpenTelemetry agent sidecar deployments and let AWS handle the infrastructure. This works across EKS and ECS, so whether you’re running Kubernetes or containers, you can push metrics without the operational overhead. It’s one of those “why didn’t they do this sooner?” features that quietly removes a whole class of deployment complexity.
https://aws.amazon.com/about-aws/whats-new/2026/07/cloudwatch-managed-collectors/
Marcin’s comment: Managed collectors for Prometheus. AWS finally read the feedback about observability being a pain, and did something about it.
That’s all for now.
Thanks again for being here, and see you in the next one.