AI News Roundup — Week of August 31, 2026

Four things happened this week that will show up on an IT budget or a risk register before they show up in a product demo: a frontier model moved inside a major CRM's trust boundary, attackers started wearing AI crawlers as a disguise, agent identity stopped being a whiteboard problem, and the way vendors bill for all of it began to change.

Claude moves inside Salesforce. Salesforce and Anthropic announced an expanded partnership on August 26 under the name Claudeforce. Per the Salesforce announcement, it runs in three directions: a Salesforce plugin inside Claude shipping with 37 prebuilt sales skills, Claude available as a reasoning model within Agentforce and its Atlas Reasoning Engine, and Claude reachable from Slack so a deal discussion can turn into a governed action without leaving the channel. The pilot is with selected customers now, with an open beta reportedly expected in September. The detail worth noting is architectural rather than commercial: Anthropic is described as the first model provider whose traffic runs entirely inside the Salesforce virtual private cloud, via Amazon Bedrock. For regulated buyers, "where does the inference actually happen" has been the blocking question on deals like this for two years.

Attackers are impersonating AI crawlers to hunt for secrets. Security firm GreyNoise reported that scanners are forging the user-agent strings of well-known AI crawlers — names belonging to OpenAI, Anthropic, Google and Perplexity among others — to sweep websites for exposed configuration files. According to GreyNoise and coverage at Help Net Security, between July 28 and August 23 six crawler identities belonging to four companies arrived on a single HTTP client fingerprint from 824 separate addresses spread across 795 distinct network ranges; a broader cluster reportedly used 13 crawler identities from eight companies. The requests targeted .env files, /.aws/credentials, private key paths and password stores. The mechanism is embarrassingly simple: a user agent is a header the client supplies, so any allowlist that checks the crawler's name but not its address can be walked straight through.

Agents get first-class identities. Okta made Agent SSO generally available on August 24, bringing its Cross App Access protocol into the main identity platform so a supported agent can be registered in the directory, assigned policy, and issued short-lived scoped tokens instead of carrying a hard-coded key. Okta says the capability is included in core SSO plans at no extra charge. The timing matches a run of standards work — NIST guidance circulating this week reportedly warns that agents authenticating with static API keys simply recreate the service-account sprawl the industry spent a decade cleaning up. Okta's own 2026 survey is the number that stings: reportedly only about a third of organisations apply the same security controls to AI agents that they apply to human staff.

Billing shifts from tokens to outcomes. CIO Dive reported on August 31 that enterprise vendors are moving toward outcome-based pricing, with OpenAI among those testing charging only when an agent completes a task successfully. The pitch to buyers is budget predictability — a finished task is a countable unit, whereas token consumption tends to reveal its true cost when the invoice lands. There is no settled standard yet; most vendors are layering outcome guarantees onto existing per-seat or consumption models rather than replacing them. Separately, the GPT-5.6 line was reportedly repriced downward across the board on August 26, continuing a price war that has been running all summer.

Our take: the crawler-impersonation story and the agent-identity story are the same story told from two ends. Both come down to a system trusting a claim it never verified — a header that says "I am Google", a token that says "I am the billing agent" — and both are fixed by the same unglamorous work: scoped, short-lived, per-identity credentials, and logs that let a human reconstruct what actually happened. If you are wiring an agent into a CRM this quarter, the Claudeforce architecture is the useful template, not the headline: know which network boundary the inference sits inside before you agree to the pilot. That is the kind of work we do — custom AI agents, cloud architecture, DevOps and security tooling — and our Odyssey Sentinel product exists on exactly this premise, that read-only, auditable visibility beats trusting a system's own account of itself. If a scan for exposed secrets across your estate sounds overdue, get in touch; it is a short engagement and it usually finds something.

Sources: Salesforce press release on Claudeforce, GreyNoise on threat actors posing as AI crawlers, Help Net Security, Okta on Agent SSO, CIO Dive on outcome-based pricing, AI Agents Directory weekly brief.