AI Update week 46

This week: Proofpoint shows how attackers use agent blueprints in Microsoft Entra to create up to 250 agent identities per blueprint in a customer tenant, Railway estimates the damage from customers’ stolen AI keys at around $740,000, Anthropic releases Claude Fable 5.2 at $8 per million input tokens, Google launches Gemini 4 Hydrogen at $1.50 per million input tokens, a breach at Fireworks AI affects less than 1 per cent of Notion workspaces, Jensen Huang steps down as CEO of NVIDIA with immediate effect and the Dutch cabinet presents its AI vision ‘Kiezen voor AI voor iedereen’ (‘Choosing AI for everyone’).

• Proofpoint shows how attackers gain access to customer tenants via agent blueprints in Entra
• Railway estimates damage from customers’ stolen AI keys at $740,000
• Anthropic releases Claude Fable 5.2 and reserves Mythos 5.2 for vetted organisations
• Google makes Gemini 4 Hydrogen the fast workhorse model alongside Carbon and Argon
• Breach at Fireworks AI exposes Notion AI chat transcripts
• Jensen Huang steps down as CEO of NVIDIA with immediate effect
• Cabinet’s AI vision prioritises sovereignty and fewer barriers for SMEs

Proofpoint shows how attackers gain access to customer tenants via agent blueprints in Entra

On 9 November Proofpoint published research into BluePhishing, a form of consent phishing in which a victim grants consent to a malicious agent blueprint from the attacker’s tenant. With that consent, the attacker can create their own agent identities in the victim’s tenant and obtain tokens for them.

• In the victim’s tenant, the blueprint is automatically granted the AgentIdentity.CreateAsManager permission, which according to Proofpoint cannot be revoked.
• Up to 250 agent identities can be linked to a single blueprint.
• According to Microsoft notice MC1506733, standard users in Entra will be able to create agent blueprints themselves from November 2026.
• Proofpoint has not yet seen any abuse in the wild and recommends measures including consent restrictions for agents, Conditional Access for agents and monitoring audit logs for new blueprints and agent identities.

Source: Proofpoint

Railway estimates damage from customers’ stolen AI keys at $740,000

On 10 November the PaaS platform Railway published its report on a breach on 20 and 21 October, in which an attacker managed to decrypt customers’ encrypted environment variables. According to Railway, unauthorised usage with stolen keys from AI providers amounts to an estimated $740,000.

• Variables may have been decrypted for up to 1.85 per cent of workspaces, around 100,000.
• Linked tokens belonging to 1,852 users were exposed, including their own keys for agents and GitHub OAuth tokens, all of which have been reset.
• At 07:28 UTC on 20 October, the attacker partially escaped the isolation of the build environment and, via a misconfigured OIDC integration, gained access to a key that could be used to decrypt variables.
• AWS Bedrock and Google Gemini are issuing refunds under their fraud policies, OpenAI is assessing claims case by case, and for OpenRouter, which does not issue refunds, Railway is covering the costs itself.

Source: Railway

Anthropic releases Claude Fable 5.2 and reserves Mythos 5.2 for vetted organisations

On 9 November Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, two variants of the same underlying model. Fable 5.2 has additional safeguards in the areas of cyber and biology and is generally available, while Mythos 5.2 lacks those safeguards and is only accessible to organisations in Anthropic’s verification programmes.

• Fable 5.2 costs $8 per million input tokens and $40 per million output tokens; the optional fast mode costs $16 and $80 and, according to Anthropic, generates tokens up to 2.5 times faster.
• According to Anthropic, Fable 5.2 scores 69.9 per cent on FrontierSWE v2, compared with 62.3 per cent for Opus 5.5 and 56.3 per cent for Fable 5.1.
• In early tests, the model blocked around 55 per cent fewer requests than Fable 5.1.
• The model is available as claude-fable-5-2 on the Claude Platform, AWS, Google Cloud and Microsoft Azure, and Claude Enterprise customers can use Claude Security, which runs on Mythos 5.2.

Source: Anthropic

Google makes Gemini 4 Hydrogen the fast workhorse model alongside Carbon and Argon

On 12 November Google introduced Gemini 4 Hydrogen, a faster and cheaper model for everyday coding work and agents. The model succeeds Gemini 3.8 Flash and is being rolled out worldwide in the API, Gemini for business, Google Sheets and Google Search.

• The model costs $1.50 per million input tokens and $7.50 per million output tokens across the full context window of 1 million tokens.
• Alongside the low, medium and high settings, there is a new xhigh thinking setting.
• According to Google, Hydrogen achieves 57 per cent on Terminal-Bench 4.1, compared with 18.2 per cent for Gemini 3.8 Flash, and completes tasks 2.5 times faster.
• According to Google, the model scores 51 on high and 54 on xhigh on the Artificial Analysis Intelligence Index, compared with 50 for Gemini 4 Argon.

Source: Google

Breach at Fireworks AI exposes Notion AI chat transcripts

On 12 November Notion published an advisory about a breach at Fireworks AI, the provider Notion used to run open-weight models. Data was affected in less than 1 per cent of Notion workspaces, mainly chat transcripts from the Notion AI Insights feature.

• Fireworks informed Notion on 2 November that malicious actors had gained access to a small set of data, after which Notion severed all integrations on 3 November.
• Notion found API keys in the affected messages: its own keys, which it revoked on 11 November, and third-party keys, which it reported to the issuers on 12 November.
• Notion has disabled the Insights feature and no longer runs open-weight models via Fireworks.
• Notion advises affected customers to treat the data as leaked and to replace any passwords and keys that were pasted into Notion AI.

Source: Notion

Jensen Huang steps down as CEO of NVIDIA with immediate effect

On 12 November NVIDIA announced that Jensen Huang has stepped down as president, CEO and board member with immediate effect. The press release gives no reason for his departure.

• CFO Colette Kress becomes interim CEO and will also remain CFO until a permanent successor is found.
• Harvey C. Jones has been appointed the first chair of the board in NVIDIA’s history.
• In its search for a successor, the board is considering both internal and external candidates.
• Results for the third quarter of fiscal 2027 will be released as scheduled on Tuesday 17 November.

Source: NVIDIA

Cabinet’s AI vision prioritises sovereignty and fewer barriers for SMEs

On 6 November the Council of Ministers approved the cabinet’s vision ‘Kiezen voor AI voor iedereen’, proposed by Willemijn Aerdts, State Secretary for Digital Economy and Sovereignty. The vision rests on three pillars: responsible development and use, access to advanced AI and security, and steering the AI transition in line with public values.

• The cabinet wants to remove obstacles that prevent SMEs from using AI.
• The cabinet wants to determine which technology and infrastructure the Netherlands and Europe need to control themselves, and will present a detailed plan for the required AI infrastructure and computing power in early 2027.
• There will be research into a Dutch AI safety institute, as well as scope to experiment with new AI applications under supervision.
• Together with the media sector, the cabinet is working on a Mediapact on the use of journalistic content by AI, and the announcement does not mention any amounts.

Source: Rijksoverheid

This article was originally published in Dutch on MSP Business. Read the original →