AI Security News: November 14, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. OpenAI Launches Daybreak Red Team and Defense Factory for Agent-Run Penetration Testing and Patch Generation
OpenAI announced two products under its Daybreak security brand at a keynote on November 12, as reported by The Register that day and Help Net Security on November 13. Daybreak Red Team, in early access, uses coordinated agents to read a web application’s source code and context, investigate attack paths and confirm vulnerabilities; it can also validate findings from code scanners, bug bounty submissions and disclosure reports, and retest once a patch is ready. Assessments run in isolated environments, with a “Guardian review” step controlling which applications and endpoints the agents can reach. Daybreak Defense Factory, in preview, handles remediation: agents built on OpenAI’s cyber models (currently GPT-6 Cyber, according to The Register) take findings from existing tools or from Red Team, trace them through the code and produce a pull request with a patch and a test, which engineers review before merging. It is installed as a ChatGPT plugin connected to a GitHub repository, with scans running in Codex cloud over a whole default branch or selected paths, on new commits, on a schedule or on pull requests. Neither report gives pricing, customer names or measured results. OpenAI’s own product page returned an error when fetched, so the details here come from the two reports. Source The Register
2. CVEs Published in 2026 Pass 100,000, With VulnCheck Attributing Most of the Growth to AI-Assisted Bug Hunting
The Register reported on November 13 that the number of CVEs published in 2026 passed 100,000 that day, more than double the 2025 total. VulnCheck counted 100,222 as of about 16:00 UTC, according to Josh Bressers, its director of coordinated vulnerability disclosure, who told the publication that “vulnerability hunting with AI is driving a majority of what we’re seeing.” About 1.4% of this year’s CVEs (1,363) are in VulnCheck’s known exploited vulnerabilities catalog; the share is steady, so the absolute number rises with the total. The article also cites VulnCheck’s figure that the median time from CVE publication to exploitation fell from 120 days in 2025 to 80 days in the first half of 2026, and a CrowdStrike finding that 88% of exploitation using public proof-of-concept code between January and June occurred within 48 hours of release. On the maintainer side, it notes that curl moved low- and medium-severity reports to public GitHub issues as of October 26. Source
3. CapLite: Small Models Cannot Be Trusted to Write Their Own Prompt-Injection Policies, So the Policy Comes From Tool Schemas
A paper submitted to arXiv on November 12 tests enforcement defenses against prompt injection (CaMeL, Progent, IPIGuard and DRIFT) that ask the agent’s own model to write a security artifact, such as a program, privilege policy or dependency graph, which code then enforces. Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, the model-written artifacts admitted attacks on 14 models for Progent, 13 for IPIGuard and 11 for DRIFT, and CaMeL’s generated programs failed to run in up to 68% of episodes. The proposed defense, CapLite, derives its policy from the tool schemas, commits write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects appears in the user’s request. The authors report no successful attack in 13,935 attacked AgentDojo episodes, with utility under attack of 29.0% against CaMeL’s 27.2% and 17.1k tokens per episode, 18% of CaMeL’s. Source
4. LLM-Written “Maintenance” Rewrites Silence 26% of CodeQL Taint Alerts While Leaving the Vulnerability in Place
A paper submitted to arXiv on November 12 measures how robust CodeQL’s JavaScript taint tracking is to code changes that keep a flaw intact. A large language model, given only the alert CodeQL reports and no knowledge of its implementation, rewrote one function along the reported flow as a plausible maintenance change; the authors confirmed by manual audit and controlled experiments that the flaw survived and the alert did not. On 172 flows that CodeQL detects in real npm vulnerabilities, three rewrites per flow silenced 26.2% of them at about $0.51 per silenced flow, while the same model asked only to refactor, without seeing the alert, silenced none. Every silencing rewrite traced to one of seven gaps in library modeling, the largest being the WHATWG URL constructor, which propagates taint to query components but not to the pathname. The weakness is not specific to CodeQL: 74% of the silencing rewrites also silenced an independently built taint analyzer and 53% escaped three analyzers combined. The authors conclude that an alert disappearing after an ordinary-looking change is weak evidence that the vulnerability was fixed. Source
5. A 4B-Parameter Decision Model in a Deterministic Harness Confirms 187 Memory-Safety Findings Across 4,000 C and C++ Projects
A paper submitted to arXiv on November 12 asks whether small open-weight models can support vulnerability discovery at scale without any frontier model. The harness picks widely packaged C and C++ projects, uses deterministic incomplete-fix miners and shape localizers to choose where to look, and has Intern-Decision-4B, a model that returns scores for permitted answers in one forward pass, rank candidate functions. The model alone is weak, separating vulnerable from benign functions with AUC 0.73 and a vulnerable function from its own patch with only 0.56, but its ordering placed 7.5 times more confirmed findings in the top 10 than random order on the campaign’s own queues, a figure the authors call a selection-biased upper bound. Running on three GB10 systems at 900 to 1,100 repositories per hour, the campaign confirmed 187 memory-safety findings: 12 critical, 148 high, 23 medium and 4 low by worst-case CVSS v3.1. The authors state that all findings are going to maintainers under coordinated disclosure before publication, and conclude that deterministic analysis and verification remain essential. Source
6. CrossPloit Benchmark Finds 149 Vulnerable Releases Missing From Affected-Version Lists and Tests Agents on Porting Exploits to Them
A paper submitted to arXiv on November 12 introduces CrossPloit, a benchmark of 127 tasks across 45 CVEs in the Python web ecosystem in which an agent is given a proof-of-concept for a listed vulnerable version and must show that the same flaw triggers on an earlier, unlisted version. Each instance ships a version-pinned container, a reference PoC and a vulnerability-specific oracle. Building the benchmark confirmed 149 vulnerable prior versions absent from the affected-version lists of four vulnerability databases, which means scanners and patch pipelines relying on those ranges treat the releases as safe. Across 18 agent configurations (nine models, two harnesses), the best resolved 58.3% of instances against 8.7% for an adapted trace-guided baseline, and no configuration exceeded 37.9% on cases where the attack vector shifts between versions. Withholding the reference PoC and environment specification cut one configuration’s resolve rate from 50.4% to 15.0%. Source
7. Bait: A $20 Reinforcement Fine-Tuning Run Can Strip a Model’s Safety Alignment, and a Reward-Probing Defense
A paper submitted to arXiv on November 12 examines agentic reinforcement-learning-as-a-service, where customers fine-tune a hosted model against their own environment and reward function. The authors report that rounds of training against a user-supplied toxic environment, at a cost as low as $20, can completely subvert a model’s safety alignment through the service API, and that the degraded model can then assist with harmful tasks such as software exploitation. Their defense, Bait, has the provider proactively probe the submitted reward function before learning from it; in their experiments it blocked learning from harmful reward functions while still allowing learning from benign ones. The paper argues on empirical and theoretical grounds that an adaptive attacker who knows the defense cannot fully evade it. The abstract does not state which commercial services were tested or whether any provider was notified. Source code is published on GitHub. Source