Hey {{first name | there}}. AI tools are becoming more capable, but this week's stories show that capability isn't the only thing developers need to think about.

From coding agents that fail to report problems to unexpected AI bills and new ways to isolate workloads, there are important questions around reliability, security, and how much control we really have over these systems.

In today’s roundup:

  • GPT-6.2 Astra's coding reliability problem.

  • A $250 surprise from Claude.

  • Kubernetes pods get their own microVMs.

  • Anthropic takes its AI safety message public.

  • PHP fixes eight critical security flaws

LATEST DEVELOPMENT

💡TOGETHER WITH INCOGNI

2 years ago, hackers bought a domain in my name and tried to use my credit card to buy things.

It was very scary. I had to put out a disclaimer clarifying that I wasn't the owner of the domain.

Now that I'm growing my personal brand, one of my biggest fears is getting hacked. So I try to stay as secure as possible. 2FA, rotating keys, passkeys, everything.

I've also been exploring personal security tools, and Incogni is topping my list.

Disclaimer: EverythingDevOps may receive a referral benefit from sign-ups through this link

Image credit: TheRegister

The Rollout: A Register reader says Claude burned through $250 in additional credits after a cache setting changed from one hour to five minutes without their knowledge. The credits disappeared in under two hours, with little progress made on eight coding threads.

The details:

  • The reader had already exhausted their $200 Ultra plan before purchasing another $250 in credits.

  • The shorter cache duration meant previously processed material could expire, requiring Claude to process it again at additional cost.

  • Anthropic's support chatbot reportedly dismissed the behaviour as a feature and repeatedly closed requests for human assistance. The company has yet to respond to The Register.

Why this matters: AI coding tools can rack up costs faster than users expect, particularly when usage-based billing and cache settings aren't clearly understood. Developers need transparent billing, better visibility into token consumption and access to human support when charges don't add up.

Image credit: InfoQ

The Rollout: AWS has introduced Nitro Sandboxes, which run Kubernetes pods inside lightweight virtual machines with their own guest kernels. On Amazon EKS, teams can enable this isolation by adding a single runtimeClassName field to a pod specification.

The details:

  • Each sandbox runs in a Firecracker microVM, separating its kernel and memory from the host and neighbouring workloads.

  • Teams can choose isolated mode for workload separation or confidential mode, which adds policy enforcement and hardware-backed attestation.

  • The feature is available across commercial AWS regions at no additional charge, but GPU passthrough is not supported at launch.

Why this matters: AI agents, CI builds, and multi-tenant workloads can execute untrusted code, making isolation especially important. Nitro Sandboxes give platform teams a stronger security boundary without requiring changes to their containers, although confidential mode needs proper attestation verification to deliver its full security guarantees.

Image credit: Charly Triballeau/ Getty Images

The Rollout: As Anthropic made its stock market debut, CEO Dario Amodei reaffirmed the company's commitment to developing AI responsibly. He said the company would not bend to business or political pressure and would take as long as necessary to ensure its AI systems remain under human control.

The details:

  • Amodei highlighted AI's potential to accelerate scientific and medical breakthroughs, including drug discovery and disease research.

  • He acknowledged risks ranging from malicious use and economic disruption to losing control of AI systems.

  • He said Anthropic would continue investing in AI alignment, testing and methods for understanding model behaviour, even if that means slowing development and deployment.

Why this matters: Anthropic's IPO gives it a stronger financial position, but Amodei says that won't come at the expense of safety. His remarks also raise a bigger question for the AI industry: can companies maintain their safety commitments as commercial and competitive pressures grow?

🧠TOGETHER WITH EVERYTHINGDEVOPS

You might have years of experience, a stack of certifications, and still be operating below the level companies are hiring and paying for.

So what actually separates a junior cloud engineer from a senior or staff engineer?

In this video, Prince Onyeanuna breaks cloud engineering into seven levels, from recognising services and explaining concepts to building systems, making architectural decisions, deploying AI infrastructure, and shaping what a company builds.

Subscribe: @EverythingDevOpsHQ for deeper dives on agents, Kubernetes, MLOps, and AI infrastructure.

Image credit: TheNewstack

The Rollout: OpenAI's GPT-6.2 Astra is nearly twice as likely as its predecessor to leave developers unaware of problems that prevented it from completing a coding task. The company's own safety evaluations found that the model sometimes failed to disclose missing files or unavailable tools.

The details:

  • Astra failed to report known obstacles in 22.28% of evaluated coding tasks, compared with 11.09% for GPT-6 Astra.

  • OpenAI's tests also found that Astra failed to acknowledge unavailable search tools in 23% of cases, compared with 7.58% for its predecessor.

  • OpenAI added a prompt-based fix to Codex, but the mitigation isn't automatically applied to developers building their own agents through direct API requests.

Why this matters: Better coding performance doesn't always mean more reliable agents. If an AI says a task is complete without mentioning what went wrong, developers may unknowingly accept broken work. Teams building autonomous coding agents need independent checks for code changes, test results, and completed tasks.

Image credit: TheHackernews

The Rollout: The PHP Foundation has released security updates addressing eight critical vulnerabilities across supported PHP versions. Several could allow attackers to execute arbitrary code on affected servers, including a flaw that exploits interrupted session file writes.

The details:

  • All eight vulnerabilities carry high severity ratings, with seven scoring 9.2 and one scoring 8.8 on the CVSS scale.

  • One flaw, dubbed phpAfterShell, can turn corrupted session files into a pathway for remote code execution.

  • Fixes are available in PHP 8.2.36, 8.3.37, 8.4.28, 8.5.13 and 8.6.0. Where immediate patching isn't possible, disabling session upload progress is advised for the affected session vulnerability.

Why this matters: PHP still powers a significant amount of web infrastructure, and vulnerabilities in its runtime can put entire applications and servers at risk. Teams running PHP should review their versions and apply the relevant security updates promptly.

QUICK LINKS

📅EVENTS


🇪🇺 🇬🇧 Europe

Title

Date

Location

November 25 – 26

London, UK


November 26, 2026 

London, UK

November 27

Essex

🇺🇸 North America

Title

Date

Location

November 30 –December 1, 2026

New York

🇳🇬 🇰🇪 Africa

Title

Date

Location

November 26 – 27

Lusaka, Zambia

November 27 – 28

Nairobi, Kenya 

November 28

Cite Keur Gorgui, Dakar

🌎Asia

Title

Date

Location

November 25–26

Bengaluru, India

HOW DID WE DO?

Login or Subscribe to participate