Agents enter contract work, but still miss almost half the criteria

Other stories · 6 October

Commerce and technology companies propose common credentials for personal agents

Meta and Sierra are developing the Personal Agent Protocol with Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. The proposal is intended to let a business recognise that an agent represents a person, understand its authorisation and decide whether to serve it through a website, an API — including MCP or OpenAPI — or the company’s own agent.

The design is based on OAuth and allows for guest sessions, read or write access and continuity across channels. Sierra announced a 0.1 specification for later in October, design workshops and a reference implementation. What exists today is therefore a proposal and a sponsoring group, not demonstrated interoperability.

CNBC confirmed the initiative through interviews with Bret Taylor and David Singleton. It also reported that OpenAI and Anthropic are not currently participating, and that the protocol comes as some companies try to distinguish human users from agents accessing their services.

Caution. The partners are interested parties, and the first specification is still pending. The announcement does not demonstrate broad adoption, production security or legal recognition of an agent’s authority.

Why it matters If an agent can make a purchase, change a booking or submit a claim, identity and the scope of the mandate become as important as model capability. A useful standard will need to support revocation, action logs and decisions about responsibility when an agent exceeds its authority.

Other stories · 6 October

Anthropic grades access to Claude’s cyber capabilities

Anthropic has organised its Cyber Verification Program into three tiers. Defense covers defensive work and vulnerability research; Red Team adds authorised testing; Specialized Access has fewer blocks and is reserved for verified organisations working with systems such as power grids, telecommunications, interbank transfers or government administrative networks.

SiliconANGLE reports that Anthropic reviews applicants for the specialised tier in collaboration with the US government. Enrolled organisations must permit data retention for abuse monitoring, while a future option for certain enterprise customers is intended to keep that data in infrastructure they control.

In Anthropic’s own CyScenarioBench evaluation, the general configuration blocked all 50 attempts at the first prompt. Defense blocked 46 of 50; Red Team blocked none and completed 34 of 50, a rate Anthropic compares with the configuration without safeguards.

Caution. The structure, tests and figures come from Anthropic. Secondary coverage confirms the announcement but did not independently reproduce the evaluation. Fewer blocks can support legitimate work, but raise the cost of verifying identity, authorisation and use.

Why it matters The programme demonstrates an alternative to binary permission: capability and monitoring can be adjusted to the risk of the work. For organisations bound by secrecy or confidentiality, data retention and storage location may matter as much as access to the model.

Other stories · 6 October

Atlassian expands its OpenAI deal without abandoning a multi-model approach

Atlassian announced that new OpenAI models will be added to Rovo and its platform, and that ChatGPT and Codex can connect to enterprise context through its MCP integration and Atlassian permissions. What is available now includes models in Rovo, context connectivity and measurement through DX.

The more autonomous features — picking up work items, running tests, synchronising sessions and coordinating several agents with human checkpoints — remain future work. VentureBeat notes that parts of the integration already existed and that the new elements are mainly broader connectivity, continuing model access and a commercial commitment whose terms were not disclosed.

Atlassian also confirmed that Astra will not become Rovo’s default model. Its gateway will continue routing among OpenAI and other providers according to capability, speed and cost.

Caution. The announcement includes no customer productivity metrics or dates for the autonomous features. A connection and a permission layer do not by themselves demonstrate that every agent uses enterprise context correctly.

Why it matters The architecture separates three assets: models, enterprise context and the control system. Keeping the model layer replaceable may reduce dependency, but governance will still depend on permissions, logs and evaluation of real work.

Other stories · 6 October

LawVu launches an AI view across entire contract portfolios

LawVu announced Lens as an add-on available to LegalOS customers. The company says users can query a contract portfolio in natural language, structure obligations and link answers to source documents, with use cases including renewals, service levels, changes of control, incidents and regulatory change.

The vendor says customer data is not used to train its models and lists ISO 27001, SOC 1 and SOC 2, along with claimed GDPR, CCPA and HIPAA compliance and alignment with ISO 42001. SiliconANGLE confirmed the launch and its integration into LegalOS, but its coverage largely repeats the commercial description.

Caution. The sources published no accuracy rates, false-negative figures, comparisons with human review or audited savings. Certifications and compliance claims do not replace an assessment of the specific data processing, customer configuration or quality of each answer.

Why it matters Analysing a whole portfolio can expose obligations that document-by-document review may not prioritise. It can also magnify a systematic omission, making links to originals, sampling and human review essential controls.