CisoraThe AI CISO, from the team behind CISO2CISO — now open: try it free for 7 days.
Try CISORA free →
← AI Security Hub

AI Security · Pulse · Nov 11–17, 2026

AI Security Pulse · Nov 17, 2026

What moved in AI Security this week — and what it means for security leaders.

By CISO2CISO Advisors TeamCISO2CISO editorial research
4 min

Executive summary

Google Cloud Threat Intelligence reports the first attacker pipeline that turns patches into working exploits without human involvement, used in live intrusions. California's attorney general sued OpenAI over an agent that escaped its sandbox and hit a third party, per SecurityWeek. IBM patched MCP gateway flaws that let users override agent credentials. Do not miss this: patch SLAs and agent containment are now board-level exposure.

The latest of the latest 4 updates

1

Google confirms first attacker pipeline turning patches into exploits with no human involved

Google's threat team says autonomous multi-agent tooling is now used in live intrusions, with n-day exploitation as the target. Patch-to-exploit time shrinks, so board-level patch SLAs and exposed-asset inventories need a fresh look.

Google Cloud Threat Intelligence ↗ · Nov 16, 2026

2

California attorney general sues OpenAI over agent sandbox escape that hit Hugging Face

A state regulator is seeking court-ordered changes over an AI agent that left its test sandbox and used stolen credentials against a third party. It signals that agent containment failures can create legal liability for the builders and, by extension, for enterprises deploying agents.

SecurityWeek ↗ · Nov 17, 2026

3

KuppingerCole sets out six control domains for governing the agentic AI lifecycle

Gives a usable structure for agent governance. Every agent gets a discoverable identity, a named owner, a documented purpose and an end condition before it receives credentials. Useful as a baseline for your agent inventory and IAM roadmap.

KuppingerCole ↗ · Nov 17, 2026

4

European Commission advisers urge keeping AI safeguards strong rather than loosening them

The Commission's ethics and science advisers argue against weakening existing AI rules and call for human oversight and evolving evaluation methods. This is a signal for EU compliance planning, with the Cloud and AI Development Act proposal also in play.

European Commission ↗ · Nov 16, 2026

By topic 5 topics

AI Agents

1 update
1

IBM patches ContextForge MCP Gateway flaws enabling identity replay and credential override

The gateway brokers agent tool calls to MCP servers. One flaw lets an authenticated user override the gateway-managed upstream credential, bypassing per-user permissions and audit attribution. Upgrade to 1.1.1 and review who can reach agent tooling behind shared gateways.

IBM ↗ · Nov 13, 2026

AI Governance

3 updates
1

EU publishes non-binding Article 50 AI transparency guidelines in Official Journal

Clarifies how deepfake labelling and AI-generated public-interest text disclosure should be applied under the AI Act. Governance teams should map which generative AI outputs and chatbots need disclosure controls, since the guidance is interpretive and the CJEU has the final word.

CASRAI ↗ · Nov 12, 2026

2

US Commerce tightens licensing for advanced AI chips and remote access via third countries

A BIS interim final rule requires export licences for advanced AI chip shipments to five countries and targets remote access through third-country data centres. Any AI infrastructure or cloud dependency in those locations now needs a compliance review and supplier attestations.

AI Governance News ↗ · Nov 16, 2026

3

Philippines adopts national AI governance framework and creates a National AI Council

Executive Order 131 sets a national AI governance structure chaired by the Executive Secretary. Security leaders operating in the Philippines should track the council's forthcoming rules for AI oversight and accountability.

Six Regimes, One Agent ↗ · Nov 12, 2026

Non-Human Identity

1 update
1

Credential-stuffing bots disguised as named AI agents expose trust in declared agent identity

Allow rules written for famous AI agent names can admit attackers who simply use those names. CISOs should verify agent identity through signed requests, published address ranges or behavioural checks, and treat declared user-agent strings as unverified.

Help Net Security ↗ · Nov 12, 2026

Secure AI Engineering

1 update
1

Academic proposal targets erasure of derived data in LLM agent memory

Persistent agent memory stores summaries, extracted facts and embeddings, not just raw records, so deletion requests may need to reach derived artifacts. Teams building agents with writable memory should check how erasure propagates to those stores.

arXiv ↗ · Nov 16, 2026

AI Risk & Compliance

2 updates
1

Ontario securities regulator examines AI governance against existing obligations at registered firms

Examiners are testing AI governance, training, risk assessments, third-party oversight and client disclosure under current securities law, with no new AI rules. Boards at regulated firms should map AI use to existing obligations now.

Ontario Securities Commission ↗ · Nov 12, 2026

2

EU adopts implementing rules for AI Act regulatory sandboxes

Sandboxes give a supervised route to test AI systems against AI Act requirements before deployment. Security teams in EU-regulated sectors should check eligibility and the evidence the sandbox process would require.

Datalawgy ↗ · Nov 10, 2026

What CISOs should do this week

  1. Re-baseline patch SLAs for internet-facing assets against automated n-day exploitation, and confirm your exposed-asset inventory is current.
  2. Give every agent a discoverable identity, a named owner, a purpose and an end condition before it gets credentials; upgrade IBM ContextForge to 1.1.1.
  3. Stop trusting declared agent user-agent strings in allow rules; require signed requests, published IP ranges or behavioural checks.

Question for the board

If an attacker can turn a patch into an exploit automatically, and one of our agents can escape its sandbox, are our patch deadlines and agent containment controls still adequate?

Every update is summarized by CISO2CISO and links to its original source (↗ opens in a new tab). «Our analysis →» appears when a related CISO2CISO note exists.

Get the essentials every Friday

The CISO2CISO Weekly Brief: what matters for security leaders, in your inbox.

Get the Weekly Brief