Three AI Operations Updates Teams Should Act On

The most important AI news is not always a new model or feature. Sometimes it is a service incident, a failed automation, or a security change that quietly affects the systems your business relies on.

On 20 July 2026, three updates deserve attention: Claude Opus 4.5 is experiencing elevated errors, GitHub Actions has recovered from a critical overnight outage, and Salesforce has started production security changes that may affect authentication and integrations.

Here is what changed, what it could mean for your workflows, and what to check next.

Status note: This article was last verified at 10:15 SAST on 20 July 2026. Service incidents can change quickly. Use the linked status pages as the current source of truth.

1. Claude Opus 4.5 is experiencing elevated errors

Anthropic is investigating elevated error rates for Claude Opus 4.5. When the incident was opened, the affected services listed included the Claude API, Claude Console, Claude Code, and Claude Cowork.

For teams using Claude in coding, research, content, support, or internal workflows, the immediate risk is not only a visible error. A failed request can also leave a longer process incomplete: a code change may not finish, an analysis may stop halfway, or a workflow may retry without producing the expected result.

What to do now

• Check the live Claude incident before repeatedly retrying a failed request.
• Record which step failed and whether any downstream action was triggered.
• If a workflow is pinned to Opus 4.5 for quality, testing, or compliance reasons, do not silently switch it to another model.
• If you use an approved fallback model, label the substitution and validate the output before treating the run as complete.

The key principle is simple: a fallback can keep work moving, but it should not quietly change the conditions under which the work was reviewed and approved.

2. GitHub Actions recovered from a critical overnight outage

GitHub Actions experienced a critical incident from 01:34 to 06:44 SAST. GitHub reported that new workflows could be delayed or fail to start, while ongoing runs could fail. The longer disruption also affected API Requests, Issues, and Pages before the platform recovered.

GitHub has marked the incident as resolved. That restores the service, but it does not prove that every workflow submitted during the incident completed correctly.

What to do now

Reconcile activity from the incident window instead of assuming automation caught up on its own:

• Review overnight Actions runs for failures, cancellations, delays, or jobs that never started.
• Confirm that expected build artefacts were created and belong to the correct commit.
• Check deployment environments and GitHub Pages for the expected release and timestamp.
• Re-run affected API, agent, scheduled, or release workflows where the result is missing or uncertain.
• Verify the final system or website, not only the workflow screen.

You can review the full timeline on the official GitHub incident page.

Recovery is a platform status. Reconciliation is a business task.

3. Salesforce production security enforcement starts today

Salesforce is starting a staggered production rollout of stronger multi-factor authentication requirements on 20 July 2026. This includes MFA enforcement for employee users and phishing-resistant MFA requirements for privileged users, including administrators, according to the applicable Salesforce release group.

At the same time, Salesforce has begun its staged global login cutover to Hyperforce. This change does not affect every Salesforce customer in the same way. The main risk is for organisations using Salesforce Express Connect, global non-My Domain login endpoints, OAuth integrations that still point to login.salesforce.com, or network allowlists based on older IP ranges.

First confirm whether Salesforce is actually in your stack

Before creating a remediation project, confirm that your business uses Salesforce in production. Check your password manager, identity provider, integration register, environment secrets, finance records, and conversations with your IT or implementation partner.

If Salesforce is not connected, record that finding and stop there. Do not create unnecessary technical work for a system you do not use.

If Salesforce is active, check the following:

• Identify employee, administrator, and other privileged accounts that fall within the MFA rollout.
• Confirm that users have an accepted MFA method and that recovery or lockout procedures are documented.
• Review Login History and integration settings for login.salesforce.com or test.salesforce.com rather than your organisation’s My Domain URL.
• Test critical SSO and OAuth connections, especially integrations that run without a person watching them.
• Review firewall and IP allowlists against Salesforce’s current Hyperforce IP data rather than relying on an old static list.
• Ask your Salesforce Express Connect provider about the migration path if private connectivity is part of your setup.

Read Salesforce’s security and MFA guidance and global login cutover guidance before changing production access controls.

The bigger lesson: AI operations need visible ownership

These three updates look different, but they point to the same operational need. Modern business systems depend on external models, code platforms, identity providers, and cloud infrastructure. A green dashboard now does not prove that an earlier job finished. A fallback model is not automatically equivalent to the model your workflow was designed around. A major security announcement does not automatically apply to every business.

A practical response has four steps:

1. Confirm whether the service is actually part of your workflow.
2. Check the official source and record the time of verification.
3. Reconcile incomplete or uncertain work.
4. Keep a human owner responsible for consequential retries, substitutions, and access changes.

That is how AI-assisted operations stay useful without becoming invisible or uncontrolled.

If your team needs a clearer place to track work, decisions, approvals, and follow-through, explore Ankor Nexa by Ankor.