A quieter day for model launches, since last week’s big releases are already covered. Today: Anthropic’s first day as a public company, two fresh agent-security advisories, a Check Point essay on why small AI guardrails fail, and a new open-weights model that ships inside Microsoft Edge.

Top 3

  • Anthropic is public. Shares began trading on Nasdaq on 19 November and closed the day about 17% above the IPO price.
  • Weak guardrails fail against strong models. Check Point argues that any filter less capable than the model behind it is eventually bypassed.
  • Edge gets an open 2.3B model. Microsoft released Aion Instruct as open weights and made it the browser’s built-in small model.

Models and platforms

Anthropic’s IPO. Anthropic priced its offering on 18 November at $83 per share, above the $70 to $78 range in the earlier filing. The stock opened at $111.50 on 19 November and closed at $97.17. Outlets disagree on the resulting market value because they count shares differently, so treat any single figure with care. The prospectus reports $17.0 billion of revenue for the quarter ended 30 September and about $900.9 billion of committed compute spending. There is no product or API change here. If you build on Claude, the useful part is the primary-source look at the vendor’s compute commitments and risk factors.

Security

Guardrails weaker than the model. On 19 November Check Point Research published an essay on “gatekeeper asymmetry doom”: if the system vetting input has less power than the system processing it, the vetting eventually fails. It ties this to its earlier PuzzleMask research, where plain-prose prompts passed four small gatekeeper models in all 23 trials while a strong target model acted on the hidden payload in 17 of 18 tests. Those tests used older models, so check your own setup. If a cheap LLM filter sits in front of a stronger agent, add output and action monitoring instead of trusting the filter alone.

ServiceNow Now Assist prompt injection. ServiceNow disclosed an indirect prompt injection in the Email Response AI Agent of Now Assist for CSM (CVE-2026-18883, rated high at 7.1 on CVSS 4.0). ServiceNow says the fix was deployed centrally, needs no upgrade, and it knows of no exploitation. The public record gives no technical detail. If you use this agent, read the vendor’s knowledge base article KB3235180 and review which actions the email agent may take.

JetBrains Junie CLI file writes. JetBrains fixed a flaw where agent file tools could modify files outside the project directory (CVE-2026-133195, 7.5 on CVSS 3.1 per JetBrains; NVD has not yet added its own score). It affects versions before 26.6.29. Update, and check your action allowlist for broad “write outside project” rules.

Agents, automation and tools

Claude Code 2.1.329. The release from the evening of 19 November adds a trust-dialog warning that says where a folder’s hooks are declared, and a warning when hooks in managed settings fail to load. It also adds scheduled tasks, /loop wakeups and forks to /tasks, and fixes ~/.claude.json sometimes being reset after a network stall or file lock. Hooks run code once you trust a folder, so update and read that dialog before accepting an unfamiliar repository.

Open source and local AI

Aion Instruct in Edge. Microsoft made Aion Instruct generally available on 19 November: a 2.3-billion-parameter model with open weights on Hugging Face. It replaces Phi-4-mini as Edge’s built-in small model and downloads only on first use, on devices that can run it well. The post also announces an origin trial of the Prompt API for live sites and a developer preview of a Decisions API that returns probabilities and a confidence value for yes/no or scoring questions without generating text. Microsoft says quality is comparable to Phi for common web tasks. If you run a web app, try the origin trial with a cloud fallback, since the model is absent on weaker devices.

Sources

Tags: , , , ,

Leave a Reply

Your email address will not be published. Required fields are marked *