cybersecurityFriday, July 3, 2026
Today's cybersecurity news is split between structural vulnerabilities in LLMs and the real-world consequences of AI-generated code. A new paper shows BPE tokenization creates exploitable gaps in safety alignment, while another researcher found a guardrail bypass in ChatGPT's file download flow. Meanwhile, Dawnguard's funding round highlights the 13x surge in AI-code vulnerabilities, and the US government confirms another breach.
Two pieces today expose fundamental weaknesses in how LLMs handle safety boundaries.
A new paper reveals that BPE tokenization creates exploitable gaps in LLM safety alignment: fragmenting safety-critical words into sub-word pieces bypasses alignment, achieving 80-100% success across five model families. This is a structural, not behavioral, vulnerability.
Researcher zer0dac found a vulnerability chain in ChatGPT combining social engineering (guardrail bypass via conversational framing) with a path traversal flaw in the file download endpoint, allowing retrieval of system files like /etc/passwd. A technical deep dive with mitigations is included.
The surge in AI-generated code vulnerabilities is drawing venture capital and scrutiny.
Dawnguard raised $6.3M to address AI-code vulnerabilities, which surged 13x in a single quarter, March 2026 alone saw more such vulnerabilities than all of 2025. The startup argues this is a mechanical consequence of AI acceleration, not careless engineering.
Nation-state threats and capacity-building efforts round out the day.
The Department of Homeland Security is investigating a cyberattack on its Homeland Security Information Network (HSIN), a platform for federal, state, and local intelligence sharing. A senior lawmaker warns the breach could risk national security.
Pakistan launched 'Cyber Secure Pakistan' to train 1 million youth in advanced cybersecurity, aiming to close the talent gap amid rising threats to government infrastructure and financial systems.
environment roundup
energy roundup
books roundup
gaming roundup
artificial intelligence roundup
programming roundup
