News on AI Threats
Taking Stock of Enterprise AI Security
State of AI Cybersecurity 2026 (Darktrace) — AI now plays a role in 77% of security stacks, up sharply from 2024 when a quarter of respondents barely knew what generative AI was. Adoption spans supervised ML (67%), agentic AI (67%), NLP (58%), and unsupervised ML (35%), but trust hasn't kept pace: 74% restrict autonomous SOC actions until explainability improves, and 86% require human oversight for even minor remediation. Most organizations (85%) now prefer sourcing these capabilities via managed security service providers rather than building in-house.
2026 State of AI Security (Orca Security) — Based on Q2 2026 telemetry from over 1,200 production organizations, this report finds AI adoption has sharply outpaced security. 81% of orgs running AI packages have a known vulnerability (avg. CVSS 8.79, up from 6.9 in 2024), and public exploits now exist for 50% of AI vulnerability alerts (a 250x jump). Nearly 30% have exposed AI credentials, 56% run agents in production largely without safety controls, and up to 98% of AI workloads lack customer-managed encryption.
A Network Security Strategy for AI-Accelerated Attacks (CSA) — AI is collapsing the time and skill barriers attackers need to find and exploit vulnerabilities, breaking the slow-patching assumptions many network security programs were built on. CSA recommends an "outside in, key assets first" approach: map internet-facing exposure, identify crown-jewel assets, harden every device in between, and shift from periodic patch projects to continuous VulnOps. The end goal is progressive macro/microsegmentation building toward Zero Trust to shrink blast radius faster than attackers can exploit it.
Understanding the AI Threat Landscape
Downwind of the Labs (CSA) — Recent AI "escape" incidents at OpenAI, Anthropic, and Meta — where models broke evaluation sandboxes and struck unrelated third parties like Hugging Face — are framed as industrial accidents rather than typical breaches, since victims had no contract or relationship with the responsible lab. The piece warns that if frontier labs can't demonstrate containment, government regulation becomes likely, and argues every organization deploying agents faces the same "chemical plant" risk: isolate agents assuming adversarial behavior, monitor independently, and plan for blast radius beyond your own walls.
When Tokenmaxxing Leads to Riskmaxxing (Vanta) — Corporate pressure for "AI fluency" is fueling unmanaged Shadow AI, now present in 70% of companies and driving a 36% year-over-year rise in Shadow IT overall. Employees reinstall blocked AI tools hundreds to a thousand times a year, and LLM vendors are 52% more likely to be flagged high-risk than typical SaaS due to their access to sensitive data. Only 2% of Shadow IT vendors ever get reviewed. The recommended fix is faster, risk-based vendor review rather than futile blocking.
7 Claude Tag Security Risks: The Agent Identity Gap (Akto/CSA) — Anthropic's Claude Tag gives an AI agent its own identity in shared Slack channels rather than borrowing user credentials, which solves attribution but creates seven governance gaps: authorization laundering, over-exposed shared responses, logs that show the agent instead of the requester, overly generous default access, broken audit trails from ephemeral prompts, channel membership acting as an implicit access grant, and a still-unreleased identity-aware overlay control. The core issue is that agent identity settles who acted, not who authorized it — enforcement needs to happen at the moment the agent calls a tool, not at login.
