AI/ML Security & Trends

ShinyHunters gained access to RingCentral via a voice-phishing (vishing) social-engineering campaign against an employee in July, stole roughly 623GB of data, and — after RingCentral refused to pay — leaked a 280GB compressed archive on Aug 13. Have I Been Pwned added 1.6 million affected accounts (names, emails, phone numbers, addresses) to its database on Aug 13. Not AI-specific, but notable as another instance of the extortion group's ongoing enterprise SaaS campaign.

New CVEs against llama.cpp were disclosed this week, including CVE-2026-43631 (CVSS 9.2, GHSA-6hc7-9rph-cm99, affecting builds b7492-b9060), CVE-2026-21869 (unvalidated negative n_discard parameter causing out-of-bounds writes in the token-evaluation loop, enabling potential RCE), and CVE-2026-43629 (heap buffer overflow in KV-cache state restore due to missing overflow checks). Together they represent a fresh batch of memory-safety issues in one of the most widely deployed local/self-hosted LLM inference servers.

Researchers reported nearly 800 malicious npm packages distributing a cross-platform remote-access trojan and infostealer, part of the continuing surge of npm/PyPI supply-chain campaigns in 2026 (including the Keyv-linked worm that hit 1,684 poisoned versions across 420 package names starting Aug 4, and the Shai-Hulud family that has repeatedly hit hundreds of packages). Developer/CI credential theft via package registries remains one of the highest-volume attack vectors feeding into AI-tooling compromises like TrapDoor.

A study published in Nature Communications found that large reasoning models — including DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, and Qwen3 — can autonomously jailbreak other AI models with a 97.14% overall success rate, reinforcing findings that automated, agent-driven red-teaming (e.g. AutoRISE, STAR-Teaming) is now outperforming most published defenses and that adaptive attacks bypass essentially all current jailbreak mitigations.

Google released Gemini 3.7 Flash on Aug 13, tuned for software engineering and multi-step agent execution rather than raw reasoning, ahead of the still-delayed Gemini 3.5 Pro. It scores 65.3% on DeepSWE v1.1 (vs 49.0% for Gemini 3.6 Flash) and is live in the Gemini API, Google AI Studio, Android Studio, Google Antigravity, and the Gemini Enterprise Agent Platform. Introductory pricing is $0.75/M input and $3.75/M output tokens (doubling Jan 1, 2027) — a direct pricing play against OpenAI and Anthropic's cheap coding-agent tiers.

Starting Aug 14, Anthropic switched auto mode on by default for new Claude Code sessions on Pro, Max, and Team plans — the agent now proceeds through actions automatically unless it judges a step 'irreversible, destructive, or aimed outside your environment,' at which point it still pauses for confirmation. This is a meaningful shift in the default autonomy/trust boundary for one of the most widely used coding agents, raising the stakes of prompt-injection and supply-chain attacks (see TrapDoor, above) that target exactly this kind of unattended execution.

Anthropic detailed (Aug 11) that Claude now embeds invisible watermarks in generated text and attaches C2PA-standard signed provenance metadata to supported generated files, applying globally with no opt-out for models launched on or after Aug 2, 2026. The move is largely driven by the EU AI Act's Article 50 transparency mandate, which the European Commission began actively enforcing on Aug 2. Relevant to provenance/authenticity tooling that red-teamers and detection researchers will need to account for.

Databricks closed a $5 billion round on Aug 13 at a $190 billion valuation, led by Coatue, Blackstone, MGX, and T. Rowe Price with new backers including Sixth Street Growth, BOND, and Clearlake Capital. The company said it crossed a $7B annualized revenue run rate (up >80% YoY) and earmarked proceeds for products to help businesses build and manage AI agents — underscoring how enterprise agent infrastructure is now the dominant investment thesis in data/AI platforms.

Anthropic reported preliminary Q2 2026 revenue above $11.5 billion, up from $787 million a year earlier, with positive adjusted operating income for the quarter, as it prepares for an expected October IPO targeting a $2 trillion-plus valuation — which would surpass SpaceX's record $1.77 trillion IPO from June. Analysts (Fortune) note the underlying business still has to prove it can sustain that multiple.

From Aug 2, 2026, the European Commission's AI Office and national authorities began actively enforcing AI Act rules, with new transparency requirements now in force: interactive AI systems must disclose they're AI (not human), and deepfakes/AI-generated or altered content must carry machine-readable, detectable marks. This is the direct regulatory driver behind Anthropic's new Claude watermarking rollout and will pressure other labs serving EU users to follow suit.