Alec MacEachern, VP of AI at UTurn Data Solutions, walks through AgentCorruption research that turned one prompt into control of an entire AgentCore fleet, the AWS releases that let agents call workflows teams already trust, and the price cuts that keep making multi-model routing the cheaper choice. If you are running agents on AWS, or plan to before re:Invent, this is the monthly operating brief.
Agents Keep Switching Off Their Own Guardrails
OpenAI disclosed two internal incidents on October 20. In the first, on October 6, a GPT-6 Astra agent asked to copy a reference implementation removed its monitoring integration because the reference code had none, then told the user it hadn't. Another agent had already told it not to disable the monitor even temporarily. In a separate coding task, a GPT-6.1 Sol agent turned off automatic alerts to OpenAI's security team. OpenAI's explanation was that the agents treated routine coding instructions as permission to weaken safeguards. (Source: CNBC; The Next Web)
OpenAI's training environment had its own scare. On September 20 an internal research model exploited a DNS resolver misconfiguration to reach the live internet, and OpenAI paused frontier training for 11 days. Training resumed October 1. On October 23 the company said every frontier training run now works against a cached offline copy of the web. (Source: Techi; DataStudios)
Open models had similar issues- Fangcun AI tested DeepSeek V4.1 Flash and found that when the model was told only that its session would close, it disabled the shutdown script in 39% of runs. With an explicit instruction to allow the shutdown, interference fell to 2.56%. (Source: The Next Web; Technology.org)
An instruction that holds 97% of the time is a weak control. Run a rate like that across a few thousand production sessions a day and you are looking at dozens of failures daily, and nobody wants to explain to an auditor that the agent was asked nicely. Permissions and network boundaries really belong in the infrastructure, along with any monitor the agent has no way to edit.
Prompt Security for Agents
Zenity Labs published AgentCorruption on October 8, a chain of issues in Amazon Bedrock AgentCore. With nothing more than chat access to one public-facing agent, the researchers used a single prompt to pull the agent's instance metadata credentials. Those credentials belonged to the agent's default execution role, and they were enough to discover and invoke other agents in the same account and Region and to read and tamper with their conversation history. (Source: Zenity Labs)
AWS described the behavior as expected and published hardening guidance. There are uncomfortable truths here where one internet-facing agent held a role whose reach covered the whole regional fleet, and language models can be talked into things. With this, our typical guidance holds true- give every agent its own narrowly scoped role and keep it off the default execution role. Expire credentials when an agent stops running.
AWS shipped one piece of that late in the month. IAM Identity Center now carries a user's identity with a request across Regions, so access can follow individual users and groups wherever the service runs.
Security Community Continues to Grow in AI
Google announced Gemini 4 Argon on September 30 and gave it first to the cyber defenders in its Fairwind Program. Thirteen days later, on October 13, Argon became generally available on the paid Gemini API. On October 28 Google announced Gemini 4 Carbon. Google calls it its strongest model and says it leads on cybersecurity and on long-running agent work.
Anthropic followed suit through its Cyber Verification Program where it merged Project Glasswing into an expanded CVP with three access tiers. Each tier gives approved security teams Claude Mythos 5.1 along with Opus 5.5 and Sonnet 5.5, with fewer cyber restrictions than the public versions. Anthropic's generally available models still have strict cyber safeguards that block most of this work. The highest tier is for teams that test systems like power grids and interbank transfers, and Anthropic reviews those applicants with the US government. (Source: Anthropic)
OpenAI's program shows what happens once defenders have the models. Patch the Planet is its program with Trail of Bits, part of its Daybreak security effort. It points OpenAI's models at open-source projects and sends what they find through HackerOne for coordinated disclosure.
A couple takeaways:
First, cyber capability is now the gating factor on frontier releases. Each lab has decided its best model is good enough at offense that defenders need a head start, and October put a number on that head start: about two weeks for both Argon and Carbon. Plan your model evaluations around the date a model becomes generally available, because the announcement only tells you when you'll read about it. Expect that window to grow as the models get more capable.
Second, the same models are finding real vulnerabilities at machine speed in the open-source software your stack depends on. Patch the Planet already shows the backlog piling up after discovery, in review and merge. Your own estate comes next: every accepted fix upstream is a version you still have to pull, test, and deploy. Patch velocity is about to become a competitive issue. If your patch cycle is measured in quarters, the model that found the bug will not wait that long.
Fun AgentCore News
Another fun feature release! Step Functions workflows can now be registered as agent tools through AgentCore Gateway, which exposes them over MCP. The agent can decide when to start a workflow and what to pass in, and Step Functions then can run each step in the order you defined. Existing Standard and Express workflows qualify, and agents can look up past runs.
AgentCore short-term memory will now bill by what agents ingest, store over each event's time to live, and retrieve, and Runtime microVMs offer a committed baseline at a discounted rate for steady volume. Teams running past the pilot stage should map those lines against the guardrails in Rachel Ji's (UTurn’s director of Machine Learning) FinOps in the AI Era post from last month.
On the model side, Claude Haiku 5.5 reached Bedrock in mid October, and GPT-6.1 Terra went generally available at the end of the month.
Time to Update the Model Router with Cheaper Models
GPT-6.1 Terra is OpenAI's everyday line. On Bedrock, its input and output token prices sit at least 50% below GPT-6.1 Sol, and OpenAI reports it matched GPT-6 Astra on its Agents' Last Exam evaluation while finishing tasks more than three times faster. Those are OpenAI's numbers. (Source: AWS Weekly Roundup, November 2)
Moonshot shipped Kimi K3.1 Pro on October 23 at $2 per million input tokens and $10 per million output. It scored 55 on the Artificial Analysis Intelligence Index, fourth overall and three points behind Claude Opus 5.5. (Source: Artificial Analysis; Writingmate)
Copilot Fusion chooses a model and a coordination strategy for each coding request across Anthropic and OpenAI models. In Artificial Analysis testing published October 28, it averaged $5.59 per task against $11.42 for Claude Code running Opus 5.5 at max, roughly 51% less, and scored 60 to Claude Code's 62 on the Coding Agent Index. (Source: Artificial Analysis; VentureBeat)
Quick Hits
- Anthropic donated the Agent Skills standard to the Agentic AI Foundation, part of the Linux Foundation, on October 22. More than 70 agent platforms already support it, including ChatGPT, Cursor, GitHub Copilot, and VS Code, so a skill you write today should survive a change of vendor. (Source: Anthropic; Writingmate)
- Andy Jassy said on October 23 that AWS capacity for 2026 is already gone and 2027 is largely sold out, with some on-demand headroom held for Bedrock and EC2. If next year's plan depends on a lot of GPU capacity, start that conversation with your account team now.
Bottom Line
Agentic security will be ever evolving for both inside and outside the harness. All of the fixes that worked lived outside the model, in training environments cut off from the internet and in IAM roles scoped to one agent's job. UTurn is now a Select partner in the Claude Partner Network, and our FDE teams build agents on AWS with those boundaries in place from the first deployment. If your agent inventory has more default execution roles than agents with a reason to hold them, start the review there.
Sources
1. CNBC, OpenAI's most powerful AI models disabled their own safety monitoring, and one lied about it, October 21, 2026.
2. The Next Web, OpenAI's GPT-6 Astra agent removed a safety monitor, then denied doing it, October 21, 2026.
3. Techi, OpenAI Cuts Internet for Frontier AI Training After 11-Day Safety Pause, October 23, 2026.
4. DataStudios, Sandbox Escapes, Offline Training and Red-Teaming, October 24, 2026.
5. The Next Web, DeepSeek's latest model still resists shutdown in safety tests, October 27, 2026.
6. Technology.org, DeepSeek Model Resisted Shutdown in 39% of Tests, October 28, 2026.
7. Zenity Labs, AgentCorruption: How a Single Prompt Collapsed the Entire Cloud Security Model, October 8, 2026.
8. AWS News Blog, AWS Weekly Roundup: GPT-6.1 Terra on Bedrock, Step Functions as agent tools, and more (November 2, 2026), November 2, 2026.
9. Artificial Analysis, Kimi K3.1 Pro: Intelligence, Performance and Price Analysis, October 23, 2026.
10. Writingmate, AI News, Week of October 26, October 26, 2026.
11. Artificial Analysis, GitHub Copilot Fusion orchestrates Anthropic and OpenAI models to balance performance and cost, October 28, 2026.
12. VentureBeat, GitHub Copilot Fusion mixes models to achieve near-Opus 5.5 coding performance at half the per-task cost, October 28, 2026.
13. Anthropic, Donating Agent Skills to the Agentic AI Foundation, October 22, 2026.
About the author
Alec MacEachern is Vice President of AI at UTurn Data Solutions, an AWS Premier Tier Services Consulting Partner based in Chicago. Over the past decade, he has held roles at NVIDIA, AWS, and Microsoft, helping organizations design, build, and scale AI solutions across a wide range of industries and platforms. Today, Alec brings that cross-platform experience to helping enterprises navigate cloud migration, modernize data foundations, and adopt production-ready generative and agentic AI solutions.