Scoped tokens for agents, Copilot model retirements, safer OpenSearch upgrades, and one log retention policy
Two AI picks: Amazon Cognito adds OAuth 2.0 token exchange so an agent can call a downstream service with a narrow token instead of the user's own, and GitHub will retire six Copilot models on December 14. Two engineering picks: OpenSearch Service upgrades that can be validated on a clone and rolled back, and account-level retention policies for CloudWatch Logs.
House Speaker Mike Johnson says there is growing consensus on limiting chip exports to China. GPU capacity plans that cross borders should allow for the rules changing.
Megaspeed is the biggest buyer of Nvidia chips in South East Asia. If you rent GPU capacity in the region, ask your provider where its hardware comes from.
This was the first of 38 Vulcan launches booked by Amazon, and the rocket is expected to be grounded. Plans that depend on Amazon Leo connectivity should allow for delay.
Microsoft said it was investigating and had no evidence of exploitation in the wild. Check which laptops and build machines run Windows 11 26H2 and watch for a fix.
Ars Technica · 13 Nov 2026
AI & Agent Engineering
Agent identityverified
Amazon Cognito adds OAuth 2.0 token exchange for services and agents acting for a user
Amazon Cognito now supports OAuth 2.0 token exchange (RFC 8693). A service or AI agent acting on a user's behalf can exchange the user's token for a new token scoped to a specific downstream service. The new token has a narrower scope, is bound to a single audience and comes without a refresh token. When an actor token is included, Cognito adds an act claim that identifies the intermediary.
Why it matters
Passing a user's original token through an agent or MCP server gives every hop the user's full access, and downstream logs cannot tell the agent from the user. The act claim lets a service tell the two apart. Token exchange is a published standard, so the design applies beyond Cognito.
Do this week
Trace one agent or MCP server call path and write down which token each hop receives. Wherever the user's original token is forwarded, plan a token exchange step, and have the downstream service log the act claim.
GitHub will deprecate six Copilot models on December 14
GitHub will deprecate six models across all Copilot experiences on December 14, 2026: Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3. Suggested alternatives are named for five of them; the one for Kimi K3 is to be announced. Alternatives are enabled automatically for Copilot Enterprise and Business unless an administrator turned off the global default or disabled the model.
Why it matters
A model named in a workflow, an integration or a team guide stops being available on that date. Organizations that turned off the global default will not get the replacements automatically, so someone has to enable them in the model policies before the cutoff.
Do this week
Search your repositories and Copilot settings for the six model names this week. Move each use to the suggested alternative, rerun any evals that depend on it, and if you disabled default model enablement, turn on the replacements.
OpenSearch Service adds clone-based upgrades that can be validated first and rolled back
Amazon OpenSearch Service now supports clone-based upgrades. The service copies data from the existing primary cluster to a cloned cluster on the new version while the primary keeps serving the workload. You can run the same search queries on both and compare them with CloudWatch metrics, then cut over. You can cancel before the cutover, or roll back afterward during a retention period you choose.
Why it matters
Search cluster upgrades are often put off because they are hard to test under real load and hard to undo. A clone on the new version turns the upgrade into a comparison followed by a reversible cutover. The same idea works for any stateful system you run yourself.
Do this week
If you run OpenSearch Service on an old version, pick one non-critical domain and write down the queries and CloudWatch metrics you would compare before a cutover. Set a retention period for the previous version and try a rollback once.
Amazon CloudWatch Logs now supports account-level retention policies. One policy sets how long log data is kept across all log groups in an account, or across those whose names start with a prefix such as /aws/lambda. You choose whether it applies to existing log groups or only new ones. When more than one policy matches, the longest retention period applies. Up to 20 policies per account and Region, at no additional cost.
Why it matters
Retention set on each log group drifts: new groups appear without it and keep data longer than anyone decided. A baseline for the account with longer periods by prefix makes retention a stated rule. Because the longest matching period wins, a baseline cannot shorten what audit needs.
Do this week
Agree a baseline with security, for example 90 days for the account and 1 year for log groups that start with /security, as in the AWS example. Apply it to new log groups first, list what it would change, then extend it to existing ones.
Find the log groups nobody set a retention period for
In one AWS account, list the CloudWatch log groups that have no retention period and note how much data each stores. Then draft an account-level retention policy: one baseline for the account and a longer period for one prefix that audit cares about. Apply it to new log groups only.
Why. It shows how much log data is kept by default, not by decision, and gives a safe first step, since a policy limited to new log groups does not touch existing data.
Measure. Record the number of log groups without a retention period, the stored bytes they hold, and how many account-level policies you needed to cover them.
Classic / essential read
Parallel Change
Danilo Sato · 5 min
Written in 2014, it describes expand and contract: support the old and new versions side by side, migrate, then remove the old one. Today's clone-based upgrade and the Copilot model switch are both changes that are safer when old and new overlap for a while.
One lesson. Break a backward-incompatible change into three phases, expand, migrate and contract, so that it can be released safely at any point.
Today's picks are about changing things safely: give an agent a narrow token instead of the user's, move off a model before it is retired, upgrade a cluster beside the old one, and set log retention once for the account instead of group by group.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.